Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠
JavaScript ai-agentai-agentsai-toolsbrowser-aibrowser-automationbrowser-extensionchrome-aichrome-devtoolschrome-extensionchrome-extensionsfirefox-addonfirefox-addonsfirefox-extensionfirefox-extensionsgemmagemma4minimaxoffline-aioffline-llmqwen
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>WebBrain Blog</title>
<meta name="description" content="Engineering notes from WebBrain - the open-source AI browser agent.">
<meta property="og:title" content="WebBrain Blog">
<meta property="og:description" content="Engineering notes from WebBrain - the open-source AI browser agent.">
<meta property="og:type" content="website">
<meta property="og:url" content="https://webbrain.one/blog">
<meta property="og:image" content="https://webbrain.one/og-image.png">
<meta property="og:image:secure_url" content="https://webbrain.one/og-image.png">
<meta property="og:image:type" content="image/png">
<meta property="og:image:width" content="1200">
<meta property="og:image:height" content="630">
<meta property="og:image:alt" content="WebBrain — Open-source AI browser agent">
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:title" content="WebBrain Blog">
<meta name="twitter:description" content="Engineering notes from WebBrain.">
<meta name="twitter:image" content="https://webbrain.one/twitter-image.png">
<meta name="twitter:image:alt" content="WebBrain — Open-source AI browser agent">
<link rel="icon" type="image/png" href="/favicon.png">
<link rel="canonical" href="https://webbrain.one/blog">
<!-- Blog is English-only; alternates point to locale homepages. -->
<link rel="alternate" hreflang="en" href="https://webbrain.one/">
<link rel="alternate" hreflang="es" href="https://webbrain.one/es/">
<link rel="alternate" hreflang="fr" href="https://webbrain.one/fr/">
<link rel="alternate" hreflang="tr" href="https://webbrain.one/tr/">
<link rel="alternate" hreflang="zh" href="https://webbrain.one/zh/">
<link rel="alternate" hreflang="ru" href="https://webbrain.one/ru/">
<link rel="alternate" hreflang="uk" href="https://webbrain.one/uk/">
<link rel="alternate" hreflang="ar" href="https://webbrain.one/ar/">
<link rel="alternate" hreflang="ja" href="https://webbrain.one/ja/">
<link rel="alternate" hreflang="ko" href="https://webbrain.one/ko/">
<link rel="alternate" hreflang="id" href="https://webbrain.one/id/">
<link rel="alternate" hreflang="th" href="https://webbrain.one/th/">
<link rel="alternate" hreflang="ms" href="https://webbrain.one/ms/">
<link rel="alternate" hreflang="tl" href="https://webbrain.one/tl/">
<link rel="alternate" hreflang="he" href="https://webbrain.one/he/">
<link rel="alternate" hreflang="x-default" href="https://webbrain.one/">
<style>
*, *::before, *::after { margin: 0; padding: 0; box-sizing: border-box; }
:root {
--bg: #0b0e17;
--bg-card: #111827;
--bg-card-hover: #1a2233;
--surface: #161d2e;
--border: rgba(255,255,255,0.07);
--text: #e4e4ec;
--text-dim: #8b8fa4;
--accent: #6c63ff;
--accent-glow: rgba(108,99,255,0.25);
--accent2: #a78bfa;
--success: #34d399;
--warning: #fbbf24;
--danger: #f87171;
--tint-soft: rgba(255,255,255,0.05);
--tint-soft-2: rgba(255,255,255,0.04);
--tint-medium: rgba(255,255,255,0.08);
--nav-bg: rgba(11,14,23,0.85);
--code-bg: #0a0e17;
--inline-code-bg: rgba(255,255,255,0.06);
--shadow-strong: rgba(0,0,0,0.40);
--radius: 12px;
--radius-lg: 16px;
--max-w: 760px;
/* Line the shared site footer up with the blog's header. */
--wb-footer-width: min(calc(var(--max-w) - 48px), calc(100% - 48px));
color-scheme: dark;
}
:root[data-theme="light"] {
--bg: #f7f1e6;
--bg-card: #fffdf8;
--bg-card-hover: #f2e9d4;
--surface: #ede2cb;
--border: rgba(89,55,25,0.15);
--text: #2c1810;
--text-dim: #6b5b47;
--accent: #5b52e8;
--accent-glow: rgba(91,82,232,0.20);
--accent2: #7c6ce6;
--success: #2d8866;
--warning: #9a6500;
--danger: #b74747;
--tint-soft: rgba(89,55,25,0.05);
--tint-soft-2: rgba(89,55,25,0.07);
--tint-medium: rgba(89,55,25,0.10);
--nav-bg: rgba(247,241,230,0.85);
--code-bg: #fff7e6;
--inline-code-bg: rgba(89,55,25,0.08);
--shadow-strong: rgba(89,55,25,0.18);
color-scheme: light;
}
html { scroll-behavior: smooth; }
body {
font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Inter, Roboto, sans-serif;
background: var(--bg);
color: var(--text);
line-height: 1.75;
font-size: 17px;
overflow-x: hidden;
}
a { color: var(--accent2); text-decoration: none; }
a:hover { text-decoration: underline; }
.glow-bg {
position: fixed;
inset: 0;
z-index: -1;
pointer-events: none;
overflow: hidden;
}
.glow-bg::before,
.glow-bg::after {
content: '';
position: absolute;
border-radius: 50%;
filter: blur(120px);
opacity: 0.26;
}
.glow-bg::before {
width: 500px;
height: 500px;
background: var(--accent);
top: -220px;
left: -160px;
}
.glow-bg::after {
width: 500px;
height: 500px;
background: var(--accent2);
right: -180px;
bottom: -230px;
opacity: 0.18;
}
nav {
position: sticky;
top: 0;
z-index: 100;
background: var(--nav-bg);
backdrop-filter: saturate(160%) blur(12px);
-webkit-backdrop-filter: saturate(160%) blur(12px);
border-bottom: 1px solid var(--border);
}
.nav-inner {
max-width: var(--max-w);
margin: 0 auto;
padding: 14px 24px;
display: flex;
align-items: center;
justify-content: space-between;
gap: 20px;
}
.nav-brand {
display: inline-flex;
align-items: center;
gap: 6px;
flex-shrink: 0;
color: var(--text);
font-size: 18px;
font-weight: 700;
text-decoration: none;
}
.nav-brand:hover { text-decoration: none; }
.nav-brand .brand-logo { width: 26px; height: 26px; border-radius: 7px; flex: 0 0 auto; }
.nav-brand .domain { opacity: 0.5; font-weight: 400; }
.nav-links {
display: flex;
align-items: center;
justify-content: flex-end;
gap: 16px;
min-width: 0;
}
.nav-links a {
position: relative;
color: var(--text-dim);
font-size: 14px;
font-weight: 500;
white-space: nowrap;
}
.nav-links a:hover { color: var(--text); text-decoration: none; }
.nav-links a[aria-current="page"] { color: var(--text); }
.nav-links a[aria-current="page"]::after {
content: '';
position: absolute;
left: 1px;
right: 1px;
bottom: -7px;
height: 3px;
border-radius: 999px;
background: linear-gradient(90deg, var(--accent2), var(--accent));
box-shadow: 0 2px 12px var(--accent-glow);
}
.lang-dropdown {
background: var(--tint-soft);
color: var(--text);
border: 1px solid var(--border);
border-radius: 8px;
padding: 6px 10px;
font-size: 13px;
max-width: 120px;
}
.lang-dropdown:focus { outline: 2px solid var(--accent-glow); outline-offset: 1px; }
.lang-dropdown option { background-color: var(--bg-card-hover); color: var(--text); }
.theme-toggle {
appearance: none;
-webkit-appearance: none;
background: var(--tint-soft-2);
border: 1px solid var(--border);
border-radius: 999px;
width: 30px;
height: 30px;
padding: 0;
display: inline-flex;
align-items: center;
justify-content: center;
flex-shrink: 0;
cursor: pointer;
color: var(--text-dim);
transition: background 0.2s, color 0.2s, border-color 0.2s, transform 0.1s;
}
.theme-toggle:hover {
background: var(--tint-medium);
border-color: var(--accent);
color: var(--text);
transform: translateY(-1px);
}
.theme-toggle:focus-visible {
outline: 2px solid var(--accent2);
outline-offset: 2px;
}
.theme-toggle svg { width: 16px; height: 16px; display: block; }
.theme-toggle .icon-sun { display: block; }
.theme-toggle .icon-moon { display: none; }
:root[data-theme="light"] .theme-toggle .icon-sun { display: none; }
:root[data-theme="light"] .theme-toggle .icon-moon { display: block; }
main,
article {
max-width: var(--max-w);
margin: 0 auto;
padding: 64px 24px 80px;
}
.sr-only {
position: absolute;
width: 1px;
height: 1px;
padding: 0;
margin: -1px;
overflow: hidden;
clip: rect(0, 0, 0, 0);
white-space: nowrap;
border: 0;
}
h1 {
font-size: 38px;
line-height: 1.2;
font-weight: 700;
letter-spacing: 0;
margin-bottom: 16px;
}
.lede {
color: var(--text-dim);
line-height: 1.6;
}
.lede {
font-size: 19px;
margin-bottom: 40px;
}
.meta {
font-size: 13px;
color: var(--text-dim);
margin-bottom: 12px;
letter-spacing: 0;
}
h2 {
font-size: 24px;
font-weight: 700;
margin-top: 48px;
margin-bottom: 14px;
letter-spacing: 0;
}
h3 {
font-size: 18px;
font-weight: 600;
margin-top: 32px;
margin-bottom: 10px;
color: var(--text);
}
h4, h5, h6 {
color: var(--text);
margin-top: 28px;
margin-bottom: 10px;
line-height: 1.35;
}
p { margin-bottom: 16px; }
ul, ol { margin: 0 0 16px 20px; }
li { margin-bottom: 6px; }
code {
font-family: 'SF Mono', 'Menlo', 'Consolas', monospace;
font-size: 0.88em;
background: var(--inline-code-bg);
padding: 2px 6px;
border-radius: 4px;
color: color-mix(in srgb, var(--accent2) 80%, var(--text));
}
pre {
background: var(--code-bg);
border: 1px solid var(--border);
border-radius: var(--radius);
padding: 18px 20px;
overflow-x: auto;
margin: 20px 0;
font-size: 14px;
line-height: 1.55;
box-shadow: 0 18px 60px var(--shadow-strong);
}
pre code {
background: transparent;
padding: 0;
color: inherit;
font-size: inherit;
}
blockquote {
border-left: 3px solid var(--accent);
padding: 4px 0 4px 18px;
margin: 20px 0;
color: var(--text-dim);
font-style: italic;
}
blockquote p:last-child { margin-bottom: 0; }
hr {
border: 0;
border-top: 1px solid var(--border);
margin: 36px 0;
}
article img {
max-width: 100%;
height: auto;
display: block;
border-radius: var(--radius);
border: 1px solid var(--border);
margin: 24px 0;
}
.table-wrap { overflow-x: auto; margin: 24px 0; }
table {
width: 100%;
min-width: 560px;
border-collapse: collapse;
font-size: 14px;
}
th, td {
text-align: left;
padding: 10px 12px;
border-bottom: 1px solid var(--border);
vertical-align: top;
}
th {
color: var(--text-dim);
font-weight: 600;
font-size: 12px;
text-transform: uppercase;
letter-spacing: 0;
background: var(--tint-soft-2);
}
td code { font-size: 12px; }
.win { color: var(--success); font-weight: 600; }
.lose { color: var(--danger); }
.meh { color: var(--warning); }
.callout {
background: color-mix(in srgb, var(--accent) 10%, transparent);
border: 1px solid color-mix(in srgb, var(--accent) 24%, transparent);
border-radius: var(--radius);
padding: 16px 20px;
margin: 24px 0;
font-size: 15px;
line-height: 1.65;
color: var(--text-dim);
}
.callout strong { color: var(--text); }
.author-box {
margin-top: 64px;
padding-top: 32px;
border-top: 1px solid var(--border);
color: var(--text-dim);
font-size: 14px;
}
.post-list { display: flex; flex-direction: column; gap: 18px; }
.post-card {
display: block;
background: var(--bg-card);
border: 1px solid var(--border);
border-radius: var(--radius-lg);
padding: 24px 26px;
transition: background 0.18s, border-color 0.18s, transform 0.18s;
}
.post-card:hover {
background: var(--bg-card-hover);
border-color: color-mix(in srgb, var(--accent) 30%, transparent);
transform: translateY(-1px);
text-decoration: none;
}
.post-meta {
font-size: 12px;
color: var(--text-dim);
letter-spacing: 0;
margin-bottom: 8px;
}
.post-title {
font-size: 21px;
font-weight: 600;
color: var(--text);
margin-bottom: 6px;
letter-spacing: 0;
line-height: 1.35;
}
.post-excerpt {
color: var(--text-dim);
font-size: 15px;
line-height: 1.6;
}
@media (max-width: 700px) {
:root { --wb-footer-width: calc(100% - 40px); }
body { font-size: 16px; }
.nav-inner { padding: 14px 20px; align-items: flex-start; }
.nav-links { gap: 12px; flex-wrap: wrap; }
.lang-dropdown { display: none; }
h1 { font-size: 28px; }
.lede { font-size: 16px; }
main, article { padding: 48px 20px 64px; }
h2 { font-size: 20px; margin-top: 36px; }
.post-card { padding: 20px; }
.post-title { font-size: 19px; }
}
@media (max-width: 420px) {
.nav-inner { gap: 12px; }
.nav-links a[href="https://github.com/webbrain-one/webbrain"] { display: none; }
.theme-toggle { width: 28px; height: 28px; }
}
</style>
<style>
[hidden] { display: none !important; }
.blog-index main { padding-top: 40px; }
.blog-tools {
position: relative;
margin: 0 0 26px;
padding-left: 18px;
scroll-margin-top: 88px;
}
.blog-tools::before {
content: '';
position: absolute;
inset: 2px auto 2px 0;
width: 2px;
border-radius: 999px;
background: linear-gradient(180deg, var(--accent2), color-mix(in srgb, var(--accent) 12%, transparent));
box-shadow: 0 0 18px var(--accent-glow);
}
.archive-search-label {
display: block;
margin-bottom: 8px;
color: var(--text-dim);
font-size: 12px;
font-weight: 600;
letter-spacing: 0.08em;
text-transform: uppercase;
}
.search-shell {
display: flex;
align-items: center;
min-height: 50px;
background: color-mix(in srgb, var(--bg-card) 92%, transparent);
border: 1px solid var(--border);
border-radius: var(--radius);
box-shadow: 0 14px 50px color-mix(in srgb, var(--shadow-strong) 58%, transparent);
transition: border-color 0.18s, box-shadow 0.18s, background 0.18s;
}
.search-shell:focus-within {
background: var(--bg-card);
border-color: color-mix(in srgb, var(--accent2) 64%, var(--border));
box-shadow: 0 0 0 3px var(--accent-glow), 0 18px 56px color-mix(in srgb, var(--shadow-strong) 64%, transparent);
}
.search-icon {
width: 18px;
height: 18px;
margin-left: 16px;
flex: 0 0 auto;
color: var(--text-dim);
}
.blog-search {
width: 100%;
min-width: 0;
padding: 12px 10px;
appearance: none;
-webkit-appearance: none;
border: 0;
outline: 0;
background: transparent;
color: var(--text);
font: inherit;
font-size: 15px;
}
.blog-search::-webkit-search-cancel-button { display: none; }
.blog-search::placeholder { color: color-mix(in srgb, var(--text-dim) 78%, transparent); }
.search-key {
margin-right: 12px;
padding: 2px 7px;
flex: 0 0 auto;
border: 1px solid var(--border);
border-bottom-color: color-mix(in srgb, var(--border) 60%, var(--text-dim));
border-radius: 5px;
background: var(--tint-soft-2);
color: var(--text-dim);
font: 11px/1.5 'SF Mono', 'Menlo', 'Consolas', monospace;
}
.search-clear {
margin-right: 8px;
padding: 6px 8px;
border: 0;
border-radius: 7px;
background: transparent;
color: var(--text-dim);
cursor: pointer;
font-size: 12px;
font-weight: 600;
}
.search-clear:hover { background: var(--tint-medium); color: var(--text); }
.search-clear:focus-visible,
.page-button:focus-visible,
.empty-clear:focus-visible,
.post-card:focus-visible {
outline: 2px solid var(--accent2);
outline-offset: 3px;
}
.archive-status {
display: flex;
align-items: center;
gap: 10px;
min-height: 24px;
margin-top: 9px;
color: var(--text-dim);
font-size: 12px;
}
.archive-status-rule {
width: 24px;
height: 1px;
background: var(--border);
}
.empty-state {
padding: 38px 24px;
border: 1px dashed color-mix(in srgb, var(--accent2) 36%, var(--border));
border-radius: var(--radius-lg);
text-align: center;
color: var(--text-dim);
background: color-mix(in srgb, var(--bg-card) 72%, transparent);
}
.empty-state strong { display: block; margin-bottom: 5px; color: var(--text); }
.empty-clear {
margin-top: 14px;
padding: 8px 12px;
border: 1px solid var(--border);
border-radius: 8px;
background: var(--tint-soft);
color: var(--text);
cursor: pointer;
font: inherit;
font-size: 13px;
font-weight: 600;
}
.empty-clear:hover { border-color: var(--accent2); background: var(--tint-medium); }
.pagination {
display: flex;
align-items: center;
justify-content: space-between;
gap: 18px;
margin-top: 28px;
padding-top: 22px;
border-top: 1px solid var(--border);
}
.page-numbers { display: flex; align-items: center; justify-content: center; gap: 6px; }
.page-ellipsis {
min-width: 24px;
color: var(--text-dim);
text-align: center;
font-size: 13px;
}
.page-button {
min-width: 36px;
height: 36px;
padding: 0 10px;
border: 1px solid var(--border);
border-radius: 8px;
background: var(--tint-soft-2);
color: var(--text-dim);
cursor: pointer;
font: inherit;
font-size: 13px;
font-weight: 600;
transition: color 0.16s, background 0.16s, border-color 0.16s, transform 0.16s;
}
.page-button:hover:not(:disabled) {
color: var(--text);
background: var(--tint-medium);
border-color: color-mix(in srgb, var(--accent2) 50%, var(--border));
transform: translateY(-1px);
}
.page-button[aria-current="page"] {
color: #fff;
border-color: var(--accent);
background: var(--accent);
box-shadow: 0 7px 22px var(--accent-glow);
}
.page-button:disabled { opacity: 0.38; cursor: not-allowed; }
.page-direction { min-width: 86px; }
@media (max-width: 700px) {
.pagination { gap: 8px; }
.page-button { min-width: 32px; height: 34px; padding: 0 8px; }
.page-direction { min-width: 72px; }
}
@media (max-width: 420px) {
.blog-tools { padding-left: 13px; }
.search-key { display: none; }
.pagination { flex-wrap: wrap; justify-content: center; }
.page-numbers { order: -1; width: 100%; }
.page-direction { flex: 1 1 40%; }
}
@media (prefers-reduced-motion: reduce) {
html { scroll-behavior: auto; }
.post-card, .page-button { transition: none; }
}
</style>
<!-- Site footer styles: start -->
<style>
.wb-footer {
--wb-footer-text: var(--text, var(--ink, #2c1810));
--wb-footer-dim: var(--text-dim, var(--muted, #6b5b47));
--wb-footer-line: var(--border, var(--line, rgba(89, 55, 25, 0.15)));
margin: 0;
padding: 0;
border-top: 1px solid var(--wb-footer-line);
background: var(--wb-footer-bg, none);
color: var(--wb-footer-dim);
font-size: 13px;
line-height: 1.5;
text-align: start;
}
.wb-footer-inner {
width: var(--wb-footer-width, min(calc(100% - 48px), 1100px));
margin: 0 auto;
padding: 26px 0 30px;
display: flex;
flex-direction: column;
gap: 18px;
}
/* Page links start on the brand's left edge; the social icons sit at
the far end of the same row. */
.wb-footer-row { display: flex; flex-wrap: wrap; align-items: center; justify-content: space-between; gap: 14px 32px; }
.wb-footer-brand { display: flex; flex-direction: column; gap: 5px; }
.wb-footer-tagline { margin: 0; color: var(--wb-footer-text); font-size: 15px; font-weight: 700; line-height: 1.4; }
.wb-footer-copy { margin: 0; }
.wb-footer-copy a { color: var(--wb-footer-text); font-weight: 650; text-decoration: none; }
.wb-footer-links { display: flex; flex-wrap: wrap; align-items: center; gap: 10px 18px; }
.wb-footer-social { display: flex; align-items: center; gap: 8px; }
.wb-footer-links a { color: var(--wb-footer-dim); font-weight: 600; text-decoration: none; }
.wb-footer a:hover { color: var(--wb-footer-text); text-decoration: none; }
.wb-footer a:focus-visible { outline: 2px solid currentColor; outline-offset: 3px; border-radius: 4px; }
.wb-footer-icon { display: inline-flex; align-items: center; justify-content: center; width: 22px; height: 22px; margin: -4px 0; border-radius: 6px; }
.wb-footer-icon:hover { background: color-mix(in srgb, currentColor 12%, transparent); }
.wb-footer-icon svg { display: block; width: 15px; height: 15px; }
/* Only shown without JavaScript, when the language picker can't run. */
.wb-footer-languages { display: none; flex-wrap: wrap; gap: 6px 12px; width: 100%; direction: ltr; }
.wb-footer-languages span { color: var(--wb-footer-text); font-weight: 700; }
@media (max-width: 860px) {
.wb-footer-inner { width: var(--wb-footer-width, calc(100% - 32px)); }
}
</style>
<!-- Site footer styles: end -->
<script>
(function () {
try {
var saved = localStorage.getItem('webbrain-theme');
var theme = (saved === 'light' || saved === 'dark')
? saved
: (window.matchMedia('(prefers-color-scheme: light)').matches ? 'light' : 'dark');
document.documentElement.setAttribute('data-theme', theme);
} catch (_) {
document.documentElement.setAttribute('data-theme', 'dark');
}
})();
</script>
<!-- Plausible analytics: start -->
<!-- Privacy-friendly analytics by Plausible -->
<script async src="https://157.230.3.9.sslip.io/js/pa-omp0F7DsOXM4sGKR50l7S.js"></script>
<script>
window.plausible=window.plausible||function(){(plausible.q=plausible.q||[]).push(arguments)},plausible.init=plausible.init||function(i){plausible.o=i||{}};
plausible.init()
</script>
<!-- Plausible analytics: end -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Blog",
"name": "WebBrain Blog",
"description": "Engineering notes from WebBrain - the open-source AI browser agent.",
"url": "https://webbrain.one/blog",
"image": "https://webbrain.one/og-image.png",
"blogPost": [
{
"@type": "BlogPosting",
"headline": "WebBrain Compass VLD v6 2B: State-of-the-Art Browser Vision and Agent Routing on Consumer Hardware",
"url": "https://webbrain.one/blog/webbrain-compass-vld-v6-2b",
"datePublished": "2026-10-31",
"description": "Comprehensive evaluation of WebBrain Compass VLD v6 2B across WebBrain's frozen 100-case browser viewport vision suite and compact agent tool routing benchmark locally on NVIDIA RTX 5060 Ti hardware.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Introducing OpenBonsai: open-source ternary quantization for tiny on-device models",
"url": "https://webbrain.one/blog/openbonsai-open-source-ternary-quantization",
"datePublished": "2026-10-29",
"description": "OpenBonsai is an open-source Apache-2.0 framework for ternary quantization, compressing 9B models like Qwen3.5-9B from 18 GB down to ~2 GB on a single RTX 5090.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Tiny Vision Models Compared III: LFM3-VLD and MiniCPM-V 4.7 Under WebBrain's Strict Browser Rubric",
"url": "https://webbrain.one/blog/tiny-vision-models-compared-iii",
"datePublished": "2026-10-29",
"description": "Official 100-case browser-vision benchmark evaluating LFM3-VLD-2B and MiniCPM-V 4.7 1B locally on dual NVIDIA RTX 5060 Ti GPUs. Format compliance, ground-truth visual comprehension, and local sub-2B latency.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Spark-X2.6, LFM3, MiniCPM-V 4.7, and WebBrain VLD Models: Compact Agent Routing on NVIDIA RTX 5060",
"url": "https://webbrain.one/blog/spark-x26-lfm3-minicpm-v47-tool-routing",
"datePublished": "2026-10-29",
"description": "Frozen compact agent benchmark comparing Spark-X2.6-4B, Liquid AI LFM3-1.7B, MiniCPM-V-4.7-1B, WebBrain-Compass-VLD-Tiny-v5, and LFM3-VLD-2B locally on dual NVIDIA RTX 5060 Ti GPUs.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Introducing Vision mode (aka \"Tesla\" mode): WebBrain on vision alone",
"url": "https://webbrain.one/blog/tesla-mode",
"datePublished": "2026-10-28",
"description": "Vision mode (aka \"Tesla\" mode) lets WebBrain act on vision alone. Like a Tesla that drives on cameras, WebBrain's Act runs see the page only as screenshots and act with the mouse and keyboard. Turn it off and WebBrain works more like Waymo, adding the page's accessibility tree.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Liquid LFM3-VL-2B vs WebBrain Compass VLD: a launch-day benchmark",
"url": "https://webbrain.one/blog/lfm3-vl-2b-vs-webbrain-compass-vld",
"datePublished": "2026-10-28",
"description": "We ran Liquid's new LFM3-VL-2B against WebBrain Compass VLD Tiny v5, our MiniCPM-V-4.7-1B fine-tune, on 4,857 BenchKit tasks per model the night LFM3 launched. LFM3 wins six of nine suites.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "The biggest challenge of developing WebBrain",
"url": "https://webbrain.one/blog/biggest-challenge-of-developing-webbrain",
"datePublished": "2026-10-19",
"description": "Model diversity and ever-changing websites: why building a browser agent that works with every LLM, big or small, across multiple browsers is WebBrain's hardest problem.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "HY4, MiMo V2.6, and Liquid LFM2.5: 600 browser-agent tests",
"url": "https://webbrain.one/blog/hy4-mimo-v26-lfm25-browser-agent-benchmark",
"datePublished": "2026-10-14",
"description": "Four new OpenRouter planners and two screenshot runs: HY4 Preview, MiMo V2.6 Pro and Flash, and Liquid LFM2.5. Pinned inputs, seventeen-model consensus, measured latency and cost, and committed raw results.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Service update: WebBrain Compass restored after upstream Novita hiccups",
"url": "https://webbrain.one/blog/service-update-webbrain-compass",
"datePublished": "2026-10-14",
"description": "A brief status update regarding temporary WebBrain Compass disruptions caused by an infrastructure issue at Novita, now resolved.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Ling 3.0 Flash VL compared with budget vision models",
"url": "https://webbrain.one/blog/ling30-flash-vl-compared-with-budget-vision-models",
"datePublished": "2026-10-07",
"description": "Ling 3.0 Flash VL scores 65 of 100 browser screenshots for $0.0054 on OpenRouter, compared with six budget Qwen vision models using the same images, prompt, and archived rubric.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Ling 3.1 Flash vs DeepSeek, HY3, GLM-5.2, and Laguna XS: free inference, costly retries",
"url": "https://webbrain.one/blog/ling31-flash-vs-deepseek-hy3-glm52-laguna-xs",
"datePublished": "2026-10-06",
"description": "Ling 3.1 Flash completes 100 pinned planner cases with 80.7% tool-family consensus, 32.8% exact consensus, 92 valid calls, and zero reported cost, after extensive Novita rate-limit recovery.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "SafeSocial is now on Google Play",
"url": "https://webbrain.one/blog/safesocial-google-play",
"datePublished": "2026-10-05",
"description": "SafeSocial is now available on Google Play: an Android app that blurs selected social-comparison triggers while you browse Instagram, with on-device image and video analysis and adjustable filters.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Ling 3.0 Flash VL vs Nex-N2.5 and MiniMax M3: a cheaper frozen planner",
"url": "https://webbrain.one/blog/ling30-flash-vl-vs-nex-n25-minimax-m3",
"datePublished": "2026-10-05",
"description": "Ling 3.0 Flash VL completes WebBrain's frozen 100-case planner test with 72% Sonnet alignment, 22 exact actions, a 1.73s median, and $0.00977 observed cost. Compare Nex-N2.5 Pro, mini, and MiniMax M3, with cache and pricing context.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Nex-N2.5 Pro vs mini: low prices meet the frozen planner test",
"url": "https://webbrain.one/blog/nex-n25-pro-vs-mini-frozen-planner",
"datePublished": "2026-10-03",
"description": "Nex-N2.5-Pro costs about $0.15 for WebBrain's frozen 100-case planner test, but trails MiniMax M3 on Sonnet alignment. Mini's OpenRouter tool route rejects the test; Qwen 3.8 27B provides separate compact context.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Spark X2.5 vs MiniCPM5: three compact browser-tool models tested",
"url": "https://webbrain.one/blog/spark-minicpm-compact-tool-routing",
"datePublished": "2026-09-26",
"description": "RTX 5090 tests of Spark-X2.5-4B, MiniCPM5-2B, and Spark-X2.5-1.7B, followed by a complete 15-model Compact comparison.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Jev and SafeSocial arrive in WebBrain 36.8.0",
"url": "https://webbrain.one/blog/webbrain-36-8-jev-safesocial",
"datePublished": "2026-09-20",
"description": "WebBrain 36.8.0 adds Jev for optional fast browser decisions and SafeSocial, our homegrown 13.6 MB local Instagram image classifier. Our results are deliberately mixed: Jev did not make our workloads faster, while the much smaller SafeSocial model works well for its narrow job.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "WebBrain Compass Tiny v2: small model, sharper tool routing",
"url": "https://webbrain.one/blog/webbrain-compass-v2",
"datePublished": "2026-09-10",
"description": "Meet WebBrain Compass Tiny v2, our MiniCPM5-2B-based tool-routing model. Compare its merged BF16 results with 11 other models, including larger reference models.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Compact tool routing compared: 10 models, from 2B to 35B",
"url": "https://webbrain.one/blog/compact-tool-routing-models-compared",
"datePublished": "2026-09-07",
"description": "An updated WebBrain Compact benchmark table comparing Compass Tiny with Gemma 4, Qwen 3.5, Qwen 3.6, Qwen 3.8, Nanbeige, and the LFM2.5 base model.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Introducing WebBrain Compass Tiny: edge-sized Compact tool routing for WebBrain",
"url": "https://webbrain.one/blog/webbrain-compass-tiny-v1",
"datePublished": "2026-09-05",
"description": "WebBrain Compass Tiny v1 is a private preview of a 2.6B model built for WebBrain's Compact tool loop. In our first routing suite it matches Qwen3.6-35B-A3B on tool-name conformity and beats the local LFM2.5 base model it was trained from by eight routing outcomes, at a size meant for edge deployment.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "DeepSeek V4 Flash Vision Exp tested: a great planner with experimental eyes",
"url": "https://webbrain.one/blog/deepseek-v4-flash-vision-openrouter",
"datePublished": "2026-08-27",
"description": "We tested DeepSeek V4 Flash Vision Exp on 200 WebBrain browser-agent cases. Its text planning beats Qwen3.8 Flash and GLM-5.3 Flash, but its experimental vision route trails both.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Qwen3.8 Flash vs GLM-5.3 Flash: the budget multimodal tier just reset",
"url": "https://webbrain.one/blog/qwen38-flash-vs-glm53-flash",
"datePublished": "2026-08-27",
"description": "We tested Qwen3.8 Flash and GLM-5.3 Flash on 400 WebBrain browser-agent cases. Both are huge multimodal MoEs with small active paths, million-token context, and near-identical list pricing. Qwen wins vision; GLM wins cost under a temporary launch discount.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Muse Glimmer",
"url": "https://webbrain.one/blog/muse-glimmer",
"datePublished": "2026-08-25",
"description": "Meta's open 30B Muse Glimmer beats every budget Qwen model in WebBrain's browser-vision benchmark, then lands above GPT-5.6 Luna on comparable first-action planner metrics—at the cost of much slower hosted inference.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Six budget Qwen vision models, compared: 600 requests, 54 cents on OpenRouter",
"url": "https://webbrain.one/blog/qwen-budget-vision-openrouter",
"datePublished": "2026-08-22",
"description": "We ran every current sub-dollar Qwen vision model on OpenRouter through WebBrain's 100-case browser-vision benchmark. Qwen3-VL-32B Instruct won on quality, the MoE 30B-A3B matched it at twice the speed, and the thinking variants billed four to five times the output tokens for little or no gain.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Tiny Vision Models Compared II: WebBrain, Gemma 4, and Larger Qwen 3.5 Models Join the Race",
"url": "https://webbrain.one/blog/tiny-vision-models-compared-ii",
"datePublished": "2026-08-22",
"description": "WebBrain VL 2 450M joins our 100-case browser-vision benchmark alongside Qwen 3.5 2B and 4B and Gemma 4 E4B, with the earlier Qwen and LFM baselines retained for context.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "From 16K to 50K Browser Screenshots: Fine-Tuning LFM2.5-VL-450M",
"url": "https://webbrain.one/blog/fine-tuning-lfm25-vl-450m-for-browser-tasks",
"datePublished": "2026-08-22",
"description": "How WebBrain turned Liquid AI's 450M vision-language model into a browser specialist using a reviewed 16,646-example first dataset, a targeted 50,000-example second dataset, LoRA, and a held-out 100-case test.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Tiny Vision Models Compared: LFM 2.5 VL vs Qwen 3.5 0.8B",
"url": "https://webbrain.one/blog/tiny-vision-models-qwen38-reference",
"datePublished": "2026-08-21",
"description": "Qwen 3.8 27B set a strong 74/100 reference in WebBrain's 100-case browser-vision benchmark. We compare it with Qwen 3.5 0.8B and Liquid AI's 450M and 1.6B local VLMs—and explain why a weak benchmark score can still describe a useful fallback.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Why WebBrain 33.0.0 is GPL",
"url": "https://webbrain.one/blog/why-webbrain-33-is-gpl",
"datePublished": "2026-08-20",
"description": "WebBrain 33.0.0 ships under GPL-3.0-or-later because the extension bundles the Xapian/libzim WebAssembly runtime for full-text Wikipedia ZIM search. Releases before 33.0.0 remain MIT.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "The world needs offline access. That's why we built Apocalypse Mode.",
"url": "https://webbrain.one/blog/why-we-built-apocalypse-mode",
"datePublished": "2026-08-17",
"description": "Disasters, wars, infrastructure failures, and the unexpected can take the internet away. WebBrain's Apocalypse Mode keeps local AI, Wikipedia, and practical references ready in the browser.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "WebBrain 31 brings offline WebGPU vision to text-only models",
"url": "https://webbrain.one/blog/webbrain-31-offline-webgpu-vision",
"datePublished": "2026-08-14",
"description": "WebBrain 31.0.0 adds optional offline vision powered by WebGPU and Liquid AI's sub-1GB LFM2.5-VL-450M, giving strong text-only models a local visual layer.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "EXL3 is the best quantization format for local AI right now — especially on Blackwell lite",
"url": "https://webbrain.one/blog/exl3-sparkinfer-blackwell-lite",
"datePublished": "2026-08-11",
"description": "EXL3 (turboderp's QTIP-derived trellis format) is the best low-bit quantization for local AI on Blackwell lite hardware. Now it runs ~3x faster through SparkInfer, the vLLM fork — including on odd GPU counts and DGX Spark.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "WebBrain MCP: Give Claude Code, OpenCode, and Codex access to your real browser session",
"url": "https://webbrain.one/blog/mcp-server-introduction",
"datePublished": "2026-08-08",
"description": "The WebBrain MCP server lets any MCP-capable coding agent — Claude Code, OpenCode, Codex, Cursor — delegate browser tasks to your real, already-authenticated browser. No headless login dances, no cookie transfers. Just ask your agent to fill out a form, check a dashboard, or read a page behind SSO, and it runs in the browser you actually use.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Qwen 3.6 27B vs Laguna XS 2.1 Vision: which local LLM is better?",
"url": "https://webbrain.one/blog/qwen36-27b-vs-laguna-xs-21-vision",
"datePublished": "2026-08-05",
"description": "Qwen 3.6 27B and Laguna XS 2.1 take radically different routes to local multimodal AI. We compare their architectures, WebBrain planner results, hardware tradeoffs, and Laguna's new experimental vision checkpoint.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "DeepSeek V4 Flash now has vision support",
"url": "https://webbrain.one/blog/deepseek-v4-flash-vision-support",
"datePublished": "2026-08-04",
"description": "WebBrain's DeepSeek V4 Flash Vision NVFP4 release adds screenshot and interface understanding to a model that was previously text-only.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Two ways to give an open model eyes: Inkling Small vs GLM-5.2 Vision",
"url": "https://webbrain.one/blog/inkling-small-glm-5-2-vision-benchmark",
"datePublished": "2026-08-02",
"description": "We ran the same frozen WebBrain UI-vision probe against Thinking Machines Inkling Small on OpenRouter and a full GLM-5.2 Vision hybrid on four RTX PRO 6000 Blackwell GPUs at RunPod.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Two AI gaps are narrowing: American open weights and Chinese frontier models",
"url": "https://webbrain.one/blog/american-chinese-open-model-frontier-gap-benchmark",
"datePublished": "2026-08-02",
"description": "A consensus-ranked, cost-aware WebBrain planner test of thirteen OpenRouter models from OpenAI, Anthropic, Google, xAI, Poolside, MiniMax, Z.ai, DeepSeek, Thinking Machines, Qwen, Moonshot, and Tencent.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Inkling makes the American open-weight comeback genuinely multimodal",
"url": "https://webbrain.one/blog/thinking-machines-inkling-openrouter-planner-benchmark",
"datePublished": "2026-07-23",
"description": "We tested Thinking Machines Inkling through OpenRouter on WebBrain's frozen 100-case browser-planner benchmark, then checked its image and audio understanding.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Nanbeige 4.2 3B punches above its size, but Qwen 3.5 9B still wins our frozen planner test",
"url": "https://webbrain.one/blog/nanbeige-42-3b-bf16-planner-benchmark",
"datePublished": "2026-07-22",
"description": "We tested Nanbeige4.2-3B in BF16 on an RTX 5090. Its official agent benchmarks beat Qwen 3.5 9B and Gemma 4 12B, but WebBrain's frozen browser-planner replay did not reproduce that lead.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Poolside Laguna M.1 reaches 73% in WebBrain, but 225B does not win",
"url": "https://webbrain.one/blog/poolside-laguna-m1-openrouter-planner-benchmark",
"datePublished": "2026-07-21",
"description": "Poolside Laguna M.1 is almost twice the size of Laguna S 2.1 and fixes much of its no-tool problem. It reached 73% Sonnet alignment in WebBrain, tying Hy3's headline score but still trailing MiniMax M3 and the strongest 27–31B local planners.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Poolside Laguna S 2.1 reaches 71% with a high-reasoning request, but still trails Hy3",
"url": "https://webbrain.one/blog/poolside-laguna-s-openrouter-planner-benchmark",
"datePublished": "2026-07-21",
"description": "Poolside's Laguna S 2.1 beats Hy3 on two published SWE benchmarks and MiniMax M3 on Terminal-Bench 2.1. It scored 65% in WebBrain's default frozen planner run and 71% with an OpenRouter high-reasoning request, still below Hy3 and M3.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "GLM-5.2 is not WebBrain's new planner reference yet",
"url": "https://webbrain.one/blog/glm52-nvidia-webbrain-planner-benchmark",
"datePublished": "2026-07-14",
"description": "We ran z-ai/glm-5.2 through WebBrain's frozen, full, mid, and compact browser-agent first-tool planner suites on NVIDIA. It should be a stronger frontier open-model candidate than the old Sonnet reference, but this narrow planner result is more cautious.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "WebMCP: Websites as Tools for AI Agents, and Why We're Excited",
"url": "https://webbrain.one/blog/webmcp-integration",
"datePublished": "2026-07-14",
"description": "WebMCP lets websites expose structured tools to AI agents. A conversation with Google's Gemini-in-Chrome team is pushing us toward integration, and we couldn't be more excited.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Poolside Laguna XS is fast and free, but not a WebBrain planner win",
"url": "https://webbrain.one/blog/poolside-laguna-xs-openrouter-planner-benchmark",
"datePublished": "2026-07-10",
"description": "We ran poolside/laguna-xs-2.1:free through WebBrain's frozen 100-case browser-agent first-tool benchmark on OpenRouter. It is very fast and free to call, but 65% Sonnet alignment, only 6 exact first actions, and malformed tool names keep it out of the planner shortlist.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "ThinkingCap Qwen 3.6 27B is a serious WebBrain planner candidate",
"url": "https://webbrain.one/blog/thinkingcap-qwen36-27b-planner-benchmark",
"datePublished": "2026-07-09",
"description": "We ran ThinkingCap-Qwen3.6-27B INT4 AutoRound through WebBrain's frozen browser-agent first-tool benchmark. The model tied Qwen 3.6 27B on all-case Sonnet alignment, returned in 2.25s median latency, and exposed a sharper boundary-behavior tradeoff.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Nex-N2-mini is cheap and fast, but not a WebBrain planner win",
"url": "https://webbrain.one/blog/nex-n2-mini-openrouter-planner-benchmark",
"datePublished": "2026-07-08",
"description": "We ran nex-agi/nex-n2-mini through WebBrain's frozen 100-case browser-agent first-tool benchmark on OpenRouter. The 35B MoE endpoint is fast and extremely cheap, but its planner quality lands well below the current shortlist.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Nemotron 3 Ultra is huge, free, and not a WebBrain planner win",
"url": "https://webbrain.one/blog/nemotron3-ultra-openrouter-planner-benchmark",
"datePublished": "2026-07-08",
"description": "We ran nvidia/nemotron-3-ultra-550b-a55b:free through WebBrain's frozen 100-case browser-agent first-tool benchmark on OpenRouter. The 550B-total MoE model completed the run, but landed well below the best planner rows.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Agents-A1 in WebBrain's frozen planner benchmark",
"url": "https://webbrain.one/blog/agents-a1-webbrain-planner-benchmark",
"datePublished": "2026-07-08",
"description": "We ran Agents-A1 AWQ INT4 through WebBrain's frozen 100-case browser-agent first-tool benchmark to test whether its agentic benchmark claims transfer to WebBrain's planner workload.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Five things you shouldn't do with WebBrain",
"url": "https://webbrain.one/blog/five-things-you-shouldnt-do-with-webbrain",
"datePublished": "2026-07-07",
"description": "A short pause from WebBrain's LLM benchmark posts: five things a browser agent can technically help with, but absolutely should not be used for.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Tencent Hy3 is a great OpenRouter planner, but text-only for now",
"url": "https://webbrain.one/blog/tencent-hy3-openrouter-planner-benchmark",
"datePublished": "2026-07-07",
"description": "We ran Tencent Hy3 through WebBrain's frozen 100-case browser-agent planner benchmark on OpenRouter. Hy3 lands in the serious hosted planner tier with 95 parsed tool calls, 20 exact first-action matches, and 73% Sonnet alignment, but the current model is still text-only.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "WebBrain now has an Ollama launch handoff",
"url": "https://webbrain.one/blog/ollama-launch-handoff",
"datePublished": "2026-07-02",
"description": "A new Ollama launch handoff can configure WebBrain for a local Ollama model from one terminal command. The integration is not upstream in Ollama yet, but it is available today from the esokullu/ollama branch while we work toward official integration.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Qwen 3.7 Plus is a serious OpenRouter rival to MiniMax M3",
"url": "https://webbrain.one/blog/qwen37-plus-openrouter-planner-benchmark",
"datePublished": "2026-07-02",
"description": "We ran qwen/qwen3.7-plus through WebBrain's frozen 100-case browser-agent planner benchmark on OpenRouter. It ties MiniMax M3 on all-case Sonnet alignment, beats it on tool-call discipline, and lands just below the Gemma 4 31B QAT, Qwen 3.6 27B, and MiniMax M2.7 cluster.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Qwen 3.6 27B NVFP4 is much faster, but not a clean planner upgrade",
"url": "https://webbrain.one/blog/qwen36-27b-nvfp4-planner-benchmark",
"datePublished": "2026-07-02",
"description": "We tested nvidia/Qwen3.6-27B-NVFP4 against WebBrain's frozen browser-agent first-tool benchmark. It is far faster than the older Qwen 3.6 27B row and enters the top 10, but loses three all-case Sonnet-alignment points on boundary cases.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Ornith-1.0-35B in WebBrain's frozen planner benchmark",
"url": "https://webbrain.one/blog/ornith-35b-webbrain-planner-benchmark",
"datePublished": "2026-06-26",
"description": "We ran Ornith-1.0-35B through WebBrain's frozen 100-case browser-agent first-tool benchmark after DeepReinforce's model card claimed wins over Qwen3.6-35B and Gemma4-31B on agentic coding benchmarks.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Tiny raw LFM 2.5 checkpoints in WebBrain's frozen planner benchmark",
"url": "https://webbrain.one/blog/tiny-lfm25-raw-planner-benchmark",
"datePublished": "2026-06-26",
"description": "We ran two tiny raw Liquid LFM 2.5 checkpoints, 230M and 350M, through WebBrain's frozen 100-case browser-agent first-tool benchmark on a local OpenAI-compatible endpoint.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "MiniMax M3 and WebBrain Compass 1.0 enter the frozen WebBrain planner benchmark",
"url": "https://webbrain.one/blog/minimax-m3-webbrain-cloud-tool-calling",
"datePublished": "2026-06-21",
"description": "We ran MiniMax M3 through OpenRouter and WebBrain Compass 1.0 through api.webbrain.one against WebBrain's frozen 100-case browser-agent first-tool benchmark, then updated the Sonnet 4.6 comparison table.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Gemma 4 12B QAT is fast enough for WebBrain, but Qwen 3.5 9B still routes a little better",
"url": "https://webbrain.one/blog/gemma-4-12b-qat-planner-benchmark",
"datePublished": "2026-06-20",
"description": "We tested google/gemma-4-12B-it-qat-w4a16-ct through vLLM against WebBrain's frozen first-tool-call browser-agent harness at concurrency 8, then ran the same vision probe used in earlier local-model posts.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Gemma 4 31B QAT quietly becomes the best local Gemma planner we have tested",
"url": "https://webbrain.one/blog/gemma-4-31b-qat-planner-benchmark",
"datePublished": "2026-06-20",
"description": "We tested google/gemma-4-31B-it-qat-w4a16-ct through vLLM against WebBrain's frozen first-tool-call browser-agent harness. It improves over the older Gemma 4 31B int4 run and narrowly edges Qwen 3.6 27B on strict first-action quality while running much faster.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "DiffusionGemma hits 0.35s median in the WebBrain local planner bench",
"url": "https://webbrain.one/blog/local-planner-q4-june-2026",
"datePublished": "2026-06-19",
"description": "DiffusionGemma, Gemma 4 12B Coder, Cohere North-Mini-Code, and VibeThinker-3B in WebBrain's frozen first-tool-call browser-agent harness.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "WebBrain Compass is live, and these are the local models we are benchmarking next",
"url": "https://webbrain.one/blog/webbrain-cloud-local-model-benchmarks",
"datePublished": "2026-06-19",
"description": "WebBrain Compass is now live in the latest main branch, so you can try WebBrain without running a local LLM or bringing your own API key. The hosted route is request-limited, but it is enough to see how the browser agent feels. Next up: benchmarks for VibeThinker-3B, Gemma 4 12B Coder, DiffusionGemma 26B, and Cohere North-Mini-Code 30B across practical local-VRAM bands.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Molmo2-8B is truly open. Our current serving path could not give it a fair browser-tool test.",
"url": "https://webbrain.one/blog/molmo2-8b-tool-calling",
"datePublished": "2026-06-12",
"description": "We tried to run Molmo2-8B through WebBrain's frozen 100-case browser-agent tool-calling benchmark. The open-source story is excellent: open weights, data, and recipe. The caveat: LM Studio's Molmo path rejected native structured tools, so the only successful run used a text-call compatibility shim.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "How WebBrain keeps a local AI agent from getting hijacked by the page it's reading",
"url": "https://webbrain.one/blog/agent-security-model",
"datePublished": "2026-06-02",
"description": "An act-mode browser agent can click, type, and submit as you — which means a malicious page that confuses the model is a real attack. Here's WebBrain's security model, the adversarial tests behind it, and what we found about model size and prompt injection.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Liquid LFM 2.5-8B-A1B on browser tool calling: where the new on-device MoE fits",
"url": "https://webbrain.one/blog/liquid-lfm25-tool-calling",
"datePublished": "2026-05-31",
"description": "We added Liquid AI's LFM 2.5-8B-A1B (8.3B total / 1.5B active MoE) to our browser-agent tool-calling benchmark. Against Claude Sonnet 4.6 it beats both small Qwens but loses to Gemma 4-E2B. Plus: a tool-schema drift bug that changes how we score everything, and why the license asymmetry (LFM's $10M revenue cliff vs Gemma 4's plain Apache 2.0) tilts the distillation call to Gemma.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "13 LLMs, 100 browser tasks, two baselines: which model actually picks the right tool?",
"url": "https://webbrain.one/blog/llm-tool-calling-benchmark",
"datePublished": "2026-05-23",
"description": "We benchmarked 13 local and API LLMs on 100 real browser-agent tool-calling tasks against both consensus voting and Claude Sonnet 4.6. The winner depends on which question you ask.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Pruning Gemma 4 26B-A4B to run on small GPUs",
"url": "https://webbrain.one/blog/gemma4-turkish-pruning",
"datePublished": "2026-05-19",
"description": "A practical Gemma 4 26B-A4B expert-pruning run: remove low-activation experts for Turkish+English workloads, then recover quality with a short LoRA heal.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Round 4: Qwen 3.5-9B-int4 punches above its weight — when 9B int4 beats 308B IQ3_S on affordance",
"url": "https://webbrain.one/blog/vision-shootout-round-4",
"datePublished": "2026-05-07",
"description": "Round 4 of the vision-model shootout for browser agents. A 9B int4 Qwen 3.5 model classifies the email-chip dropdown explicitly — something the 308B MiMo IQ3_S and the larger Qwen 3.6-A3B both missed. Cheapest image tokens after Gemma. The catch: it loses the red-border visual cue and joins the 'Unknowns: None' club.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Round 3: Xiaomi MiMo V2.5 enters the vision shootout — and joins the \"Unknowns: None\" club",
"url": "https://webbrain.one/blog/vision-shootout-round-3",
"datePublished": "2026-05-07",
"description": "A third round of vision-model benchmarking for browser agents. Xiaomi's MiMo-V2.5 — the \"looks promising\" model from our research notes — sits down at the same probe with the same Google sign-in screen. Strong OCR, expensive image tokens, no calibrated uncertainty.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Xiaomi MiMo V2.5 Pro vs \"V2.5 Flash\": should WebBrain add both?",
"url": "https://webbrain.one/blog/mimo-v25-pro-vs-flash",
"datePublished": "2026-04-30",
"description": "Research notes on Xiaomi's MiMo V2.5 Pro and the V2.5/Flash-style tier: why these multimodal models look promising for WebBrain compared with text-only alternatives, and how they slot into our routing policy alongside Qwen 3.6.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Round 2: Nemotron Omni 30B vs Qwen 3.6 — does cheaper image tokens beat calibrated uncertainty?",
"url": "https://webbrain.one/blog/vision-shootout-round-2",
"datePublished": "2026-04-29",
"description": "A second round of vision-model benchmarking for browser agents. NVIDIA's Nemotron Omni 30B-A3B-Reasoning and the dense Qwen 3.6-27B both enter — and one of them is 17% cheaper per image. We measure what changes for a real browser-agent workload, and what doesn't.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
},
{
"@type": "BlogPosting",
"headline": "Four vision models, one screenshot: which one is actually worth running locally for a browser agent?",
"url": "https://webbrain.one/blog/vision-model-shootout",
"datePublished": "2026-04-21",
"description": "We benchmarked Gemma 4-E2B, Gemma 4-31B, Qwen3.5-27B, and Qwen3.6-35B-A3B on the same browser screenshot using WebBrain's exact vision-sub-call prompt. Here's what the numbers look like and why one model changed our default recommendation.",
"author": {
"@type": "Person",
"name": "Emre Sokullu",
"url": "https://emresokullu.com"
}
}
]
}
</script>
</head>
<body class="blog-index">
<div class="glow-bg"></div>
<nav>
<div class="nav-inner">
<a href="/" class="nav-brand"><img class="brand-logo" src="/logo-github.png" alt="" aria-hidden="true"> WebBrain<span class="domain">.one</span></a>
<div class="nav-links">
<a href="/">Home</a>
<a href="/docs/">Docs</a>
<a href="/blog" aria-current="page">Blog</a>
<select class="lang-dropdown" id="lang-dropdown" aria-label="Select language" data-current="en"><option value="en">English</option><option value="es">Español</option><option value="fr">Français</option><option value="tr">Türkçe</option><option value="zh">中文</option><option value="ru">Русский</option><option value="uk">Українська</option><option value="ar">العربية</option><option value="ja">日本語</option><option value="ko">한국어</option><option value="id">Bahasa Indonesia</option><option value="th">ไทย</option><option value="ms">Bahasa Melayu</option><option value="tl">Filipino</option><option value="he">עברית</option></select>
<button class="theme-toggle" type="button" id="theme-toggle" aria-label="Toggle light / dark theme" title="Toggle light / dark theme">
<svg class="icon-sun" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true">
<circle cx="12" cy="12" r="4"/>
<path d="M12 2v2M12 20v2M4.93 4.93l1.41 1.41M17.66 17.66l1.41 1.41M2 12h2M20 12h2M4.93 19.07l1.41-1.41M17.66 6.34l1.41-1.41"/>
</svg>
<svg class="icon-moon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true">
<path d="M21 12.79A9 9 0 1 1 11.21 3 7 7 0 0 0 21 12.79z"/>
</svg>
</button>
<a href="https://github.com/webbrain-one/webbrain" target="_blank" rel="noopener">GitHub</a>
</div>
</div>
</nav>
<main>
<h1 class="sr-only">WebBrain Blog</h1>
<section class="blog-tools" id="blog-tools" aria-label="Search the blog archive">
<label class="archive-search-label" for="blog-search">Search the archive</label>
<div class="search-shell">
<svg class="search-icon" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" aria-hidden="true">
<circle cx="11" cy="11" r="7"></circle>
<path d="m20 20-4-4"></path>
</svg>
<input class="blog-search" id="blog-search" type="search" placeholder="Models, benchmarks, tools…" autocomplete="off" spellcheck="false" aria-controls="post-list">
<button class="search-clear" id="search-clear" type="button" hidden>Clear</button>
<kbd class="search-key" aria-hidden="true">/</kbd>
</div>
<div class="archive-status" role="status" aria-live="polite" aria-atomic="true">
<span id="results-count">68 notes</span>
<span class="archive-status-rule" aria-hidden="true"></span>
<span id="page-status"></span>
</div>
</section>
<div class="post-list" id="post-list">
<a href="/blog/webbrain-compass-vld-v6-2b" class="post-card" data-search="WebBrain Compass VLD v6 2B: State-of-the-Art Browser Vision and Agent Routing on Consumer Hardware WebBrain Compass VLD v6 2B: State-of-the-Art Browser Vision and Agent Routing on Consumer Hardware WebBrain Compass VLD v6 2B sets a new all-time state-of-the-art on WebBrain's browser vision benchmark with 66/100 strict passes and a 90.2% rubric score, alongside #1 compact tool routing at 42.7% loose match with zero anti-patterns. Comprehensive evaluation of WebBrain Compass VLD v6 2B across WebBrain's frozen 100-case browser viewport vision suite and compact agent tool routing benchmark locally on NVIDIA RTX 5060 Ti hardware. October 31, 2026 2026-10-31 WebBrain Compass LFM3-VLD Liquid AI browser vision vision benchmark tool routing agent benchmark NVIDIA RTX 5060 Ti llama.cpp WebBrain multimodal on-device AI">
<div class="post-meta">October 31, 2026 · 8 min read</div>
<div class="post-title">WebBrain Compass VLD v6 2B: State-of-the-Art Browser Vision and Agent Routing on Consumer Hardware</div>
<div class="post-excerpt">WebBrain Compass VLD v6 2B sets a new all-time state-of-the-art on WebBrain's browser vision benchmark with 66/100 strict passes and a 90.2% rubric score, alongside #1 compact tool routing at 42.7% loose match with zero anti-patterns.</div>
</a>
<a href="/blog/openbonsai-open-source-ternary-quantization" class="post-card" data-search="Introducing OpenBonsai: open-source ternary quantization for tiny on-device models Introducing OpenBonsai: open-source ternary quantization for tiny on-device models Prism ML showed that 9B models can fit into 2 GB with ternary weights, but the compression pipeline stayed closed. OpenBonsai is an Apache-2.0 recipe to ternarize your own models and fine-tunes. OpenBonsai is an open-source Apache-2.0 framework for ternary quantization, compressing 9B models like Qwen3.5-9B from 18 GB down to ~2 GB on a single RTX 5090. October 29, 2026 2026-10-29 OpenBonsai ternary quantization tiny models on-device AI WebBrain Qwen3.5-9B Prism ML Ternary Bonsai llama.cpp local AI">
<div class="post-meta">October 29, 2026 · 4 min read</div>
<div class="post-title">Introducing OpenBonsai: open-source ternary quantization for tiny on-device models</div>
<div class="post-excerpt">Prism ML showed that 9B models can fit into 2 GB with ternary weights, but the compression pipeline stayed closed. OpenBonsai is an Apache-2.0 recipe to ternarize your own models and fine-tunes.</div>
</a>
<a href="/blog/tiny-vision-models-compared-iii" class="post-card" data-search="Tiny Vision Models Compared III: LFM3-VLD and MiniCPM-V 4.7 Under WebBrain's Strict Browser Rubric Tiny Vision Models Compared III: LFM3-VLD and MiniCPM-V 4.7 Under WebBrain's Strict Browser Rubric Testing LFM3-VLD-2B and MiniCPM-V 4.7 1B on WebBrain's frozen 100-case browser viewport vision suite reveals a tale of two models: strict formatting discipline versus raw visual perception. Official 100-case browser-vision benchmark evaluating LFM3-VLD-2B and MiniCPM-V 4.7 1B locally on dual NVIDIA RTX 5060 Ti GPUs. Format compliance, ground-truth visual comprehension, and local sub-2B latency. October 29, 2026 2026-10-29 MiniCPM-V 4.7 LFM3-VLD Liquid AI OpenBMB browser vision vision benchmark NVIDIA RTX 5060 Ti llama.cpp WebBrain multimodal on-device AI">
<div class="post-meta">October 29, 2026 · 9 min read</div>
<div class="post-title">Tiny Vision Models Compared III: LFM3-VLD and MiniCPM-V 4.7 Under WebBrain's Strict Browser Rubric</div>
<div class="post-excerpt">Testing LFM3-VLD-2B and MiniCPM-V 4.7 1B on WebBrain's frozen 100-case browser viewport vision suite reveals a tale of two models: strict formatting discipline versus raw visual perception.</div>
</a>
<a href="/blog/spark-x26-lfm3-minicpm-v47-tool-routing" class="post-card" data-search="Spark-X2.6, LFM3, MiniCPM-V 4.7, and WebBrain VLD Models: Compact Agent Routing on NVIDIA RTX 5060 Spark-X2.6, LFM3, MiniCPM-V 4.7, and WebBrain VLD Models: Compact Agent Routing on NVIDIA RTX 5060 Official Q4 GGUF runs of Spark-X2.6-4B, LFM3-1.7B, MiniCPM-V 4.7, WebBrain-Compass-VLD-Tiny-v5, and LFM3-VLD-2B on WebBrain's frozen browser benchmark, featuring parallel execution across dual RTX 5060 Ti GPUs and an expanded 20-model historical leaderboard. Frozen compact agent benchmark comparing Spark-X2.6-4B, Liquid AI LFM3-1.7B, MiniCPM-V-4.7-1B, WebBrain-Compass-VLD-Tiny-v5, and LFM3-VLD-2B locally on dual NVIDIA RTX 5060 Ti GPUs. October 29, 2026 2026-10-29 Spark-X2.6 Liquid AI LFM3 MiniCPM-V 4.7 WebBrain Compass VLD LFM3-VLD lfm3-vld-2b webbrain-compass-vld-tiny-v5 browser agent benchmark tool calling NVIDIA RTX 5060 Ti llama.cpp">
<div class="post-meta">October 29, 2026 · 9 min read</div>
<div class="post-title">Spark-X2.6, LFM3, MiniCPM-V 4.7, and WebBrain VLD Models: Compact Agent Routing on NVIDIA RTX 5060</div>
<div class="post-excerpt">Official Q4 GGUF runs of Spark-X2.6-4B, LFM3-1.7B, MiniCPM-V 4.7, WebBrain-Compass-VLD-Tiny-v5, and LFM3-VLD-2B on WebBrain's frozen browser benchmark, featuring parallel execution across dual RTX 5060 Ti GPUs and an expanded 20-model historical leaderboard.</div>
</a>
<a href="/blog/tesla-mode" class="post-card" data-search="Introducing Vision mode (aka "Tesla" mode): WebBrain on vision alone Introducing Vision mode (aka "Tesla" mode): WebBrain on vision alone Vision mode (aka "Tesla" mode) lets Act runs see the page through screenshots, like a car that drives on cameras alone. Turn it off and WebBrain works more like Waymo, reading the page's structure too. Vision mode (aka "Tesla" mode) lets WebBrain act on vision alone. Like a Tesla that drives on cameras, WebBrain's Act runs see the page only as screenshots and act with the mouse and keyboard. Turn it off and WebBrain works more like Waymo, adding the page's accessibility tree. October 28, 2026 2026-10-28 WebBrain Tesla mode vision mode browser agent computer use screenshots accessibility tree vision language model">
<div class="post-meta">October 28, 2026 · 3 min read</div>
<div class="post-title">Introducing Vision mode (aka "Tesla" mode): WebBrain on vision alone</div>
<div class="post-excerpt">Vision mode (aka "Tesla" mode) lets Act runs see the page through screenshots, like a car that drives on cameras alone. Turn it off and WebBrain works more like Waymo, reading the page's structure too.</div>
</a>
<a href="/blog/lfm3-vl-2b-vs-webbrain-compass-vld" class="post-card" data-search="Liquid LFM3-VL-2B vs WebBrain Compass VLD: a launch-day benchmark Liquid LFM3-VL-2B vs WebBrain Compass VLD: a launch-day benchmark LFM3-VL-2B beats Compass VLD Tiny v5 on six of nine general vision and text suites, ties three, and loses none. Compass is still the only one of the two that is also a decision model. Next: an LFM3-based VLD. We ran Liquid's new LFM3-VL-2B against WebBrain Compass VLD Tiny v5, our MiniCPM-V-4.7-1B fine-tune, on 4,857 BenchKit tasks per model the night LFM3 launched. LFM3 wins six of nine suites. October 28, 2026 2026-10-28 Liquid LFM3 LFM3-VL-2B d1-2B WebBrain Compass VLD Tiny v5 MiniCPM-V-4.7-1B BenchKit vision language model decision model llama.cpp RunPod">
<div class="post-meta">October 28, 2026 · 8 min read</div>
<div class="post-title">Liquid LFM3-VL-2B vs WebBrain Compass VLD: a launch-day benchmark</div>
<div class="post-excerpt">LFM3-VL-2B beats Compass VLD Tiny v5 on six of nine general vision and text suites, ties three, and loses none. Compass is still the only one of the two that is also a decision model. Next: an LFM3-based VLD.</div>
</a>
<a href="/blog/biggest-challenge-of-developing-webbrain" class="post-card" data-search="The biggest challenge of developing WebBrain The biggest challenge of developing WebBrain We get asked about our hardest problem often. The answer is twofold: supporting every model — tiny ones included — and coping with websites that never stop changing. Model diversity and ever-changing websites: why building a browser agent that works with every LLM, big or small, across multiple browsers is WebBrain's hardest problem. October 19, 2026 2026-10-19 WebBrain browser agent model diversity tiny models LLM quirks cross-browser web automation challenge">
<div class="post-meta">October 19, 2026 · 2 min read</div>
<div class="post-title">The biggest challenge of developing WebBrain</div>
<div class="post-excerpt">We get asked about our hardest problem often. The answer is twofold: supporting every model — tiny ones included — and coping with websites that never stop changing.</div>
</a>
<a href="/blog/hy4-mimo-v26-lfm25-browser-agent-benchmark" class="post-card" data-search="HY4, MiMo V2.6, and Liquid LFM2.5: 600 browser-agent tests HY4, MiMo V2.6, and Liquid LFM2.5: 600 browser-agent tests MiMo Pro leads the four new planners on consensus and ideal-tool choices. Flash reaches 81 screenshot passes with a 1.20-second planner median and a one-cent planner bill. HY4 and Liquid trail on exact-action agreement. Four new OpenRouter planners and two screenshot runs: HY4 Preview, MiMo V2.6 Pro and Flash, and Liquid LFM2.5. Pinned inputs, seventeen-model consensus, measured latency and cost, and committed raw results. October 14, 2026 2026-10-14 Tencent HY4 Preview MiMo V2.6 Pro MiMo V2.6 Flash Liquid LFM2.5-24B-A2B OpenRouter browser agent benchmark tool calling vision model open weights">
<div class="post-meta">October 14, 2026 · 9 min read</div>
<div class="post-title">HY4, MiMo V2.6, and Liquid LFM2.5: 600 browser-agent tests</div>
<div class="post-excerpt">MiMo Pro leads the four new planners on consensus and ideal-tool choices. Flash reaches 81 screenshot passes with a 1.20-second planner median and a one-cent planner bill. HY4 and Liquid trail on exact-action agreement.</div>
</a>
<a href="/blog/service-update-webbrain-compass" class="post-card" data-search="Service update: WebBrain Compass restored after upstream Novita hiccups Service update: WebBrain Compass restored after upstream Novita hiccups Intermittent errors affecting WebBrain Compass have been resolved following an upstream infrastructure issue with Novita. Service is back to normal. A brief status update regarding temporary WebBrain Compass disruptions caused by an infrastructure issue at Novita, now resolved. October 14, 2026 2026-10-14 WebBrain Compass service update Novita system status">
<div class="post-meta">October 14, 2026 · 2 min read</div>
<div class="post-title">Service update: WebBrain Compass restored after upstream Novita hiccups</div>
<div class="post-excerpt">Intermittent errors affecting WebBrain Compass have been resolved following an upstream infrastructure issue with Novita. Service is back to normal.</div>
</a>
<a href="/blog/ling30-flash-vl-compared-with-budget-vision-models" class="post-card" data-search="Ling 3.0 Flash VL compared with budget vision models Ling 3.0 Flash VL compared with budget vision models Ling costs less than every budget Qwen reference in our saved runs, with 65 strict passes and 3.57-second mean latency. Reasoning still appears despite the disabling kwargs, and 22 responses hit the output limit. Ling 3.0 Flash VL scores 65 of 100 browser screenshots for $0.0054 on OpenRouter, compared with six budget Qwen vision models using the same images, prompt, and archived rubric. October 7, 2026 2026-10-07 Ling 3.0 Flash VL inclusionAI budget vision models Qwen3-VL Qwen3.5-35B-A3B OpenRouter screenshot benchmark browser agent">
<div class="post-meta">October 7, 2026 · 7 min read</div>
<div class="post-title">Ling 3.0 Flash VL compared with budget vision models</div>
<div class="post-excerpt">Ling costs less than every budget Qwen reference in our saved runs, with 65 strict passes and 3.57-second mean latency. Reasoning still appears despite the disabling kwargs, and 22 responses hit the output limit.</div>
</a>
<a href="/blog/ling31-flash-vs-deepseek-hy3-glm52-laguna-xs" class="post-card" data-search="Ling 3.1 Flash vs DeepSeek, HY3, GLM-5.2, and Laguna XS: free inference, costly retries Ling 3.1 Flash vs DeepSeek, HY3, GLM-5.2, and Laguna XS: free inference, costly retries Ling leads tool-family agreement and beats the four text-only references on ideal-tool count, but trails on exact-action consensus. Its free Novita route required 298 rate-limit responses across the benchmark and diagnostic campaign. Ling 3.1 Flash completes 100 pinned planner cases with 80.7% tool-family consensus, 32.8% exact consensus, 92 valid calls, and zero reported cost, after extensive Novita rate-limit recovery. October 6, 2026 2026-10-06 Ling 3.1 Flash inclusionAI DeepSeek V4 Flash Tencent HY3 GLM-5.2 Poolside Laguna XS 2.1 Ling 3.0 Flash VL Nex-N2.5-Pro Nex-N2.5-mini MiniMax M3 OpenRouter browser agent benchmark tool calling">
<div class="post-meta">October 6, 2026 · 8 min read</div>
<div class="post-title">Ling 3.1 Flash vs DeepSeek, HY3, GLM-5.2, and Laguna XS: free inference, costly retries</div>
<div class="post-excerpt">Ling leads tool-family agreement and beats the four text-only references on ideal-tool count, but trails on exact-action consensus. Its free Novita route required 298 rate-limit responses across the benchmark and diagnostic campaign.</div>
</a>
<a href="/blog/safesocial-google-play" class="post-card" data-search="SafeSocial is now on Google Play SafeSocial is now on Google Play SafeSocial is now available on Google Play. The Android app helps soften selected social-comparison triggers in Instagram with on-device filtering and settings you control. SafeSocial is now available on Google Play: an Android app that blurs selected social-comparison triggers while you browse Instagram, with on-device image and video analysis and adjustable filters. October 5, 2026 2026-10-05 SafeSocial Google Play Android Instagram social comparison on-device AI WebBrain">
<div class="post-meta">October 5, 2026 · 4 min read</div>
<div class="post-title">SafeSocial is now on Google Play</div>
<div class="post-excerpt">SafeSocial is now available on Google Play. The Android app helps soften selected social-comparison triggers in Instagram with on-device filtering and settings you control.</div>
</a>
<a href="/blog/ling30-flash-vl-vs-nex-n25-minimax-m3" class="post-card" data-search="Ling 3.0 Flash VL vs Nex-N2.5 and MiniMax M3: a cheaper frozen planner Ling 3.0 Flash VL vs Nex-N2.5 and MiniMax M3: a cheaper frozen planner Ling doubles Nex Pro's exact-action count and runs much faster. MiniMax M3 keeps the stronger Sonnet alignment. Ling's sub-cent replay cost benefits from a promotional route and 94.7% cached prompt tokens. Ling 3.0 Flash VL completes WebBrain's frozen 100-case planner test with 72% Sonnet alignment, 22 exact actions, a 1.73s median, and $0.00977 observed cost. Compare Nex-N2.5 Pro, mini, and MiniMax M3, with cache and pricing context. October 5, 2026 2026-10-05 Ling 3.0 Flash VL inclusionAI Nex-N2.5-Pro Nex-N2.5-mini MiniMax M3 OpenRouter browser agent benchmark tool calling">
<div class="post-meta">October 5, 2026 · 7 min read</div>
<div class="post-title">Ling 3.0 Flash VL vs Nex-N2.5 and MiniMax M3: a cheaper frozen planner</div>
<div class="post-excerpt">Ling doubles Nex Pro's exact-action count and runs much faster. MiniMax M3 keeps the stronger Sonnet alignment. Ling's sub-cent replay cost benefits from a promotional route and 94.7% cached prompt tokens.</div>
</a>
<a href="/blog/nex-n25-pro-vs-mini-frozen-planner" class="post-card" data-search="Nex-N2.5 Pro vs mini: low prices meet the frozen planner test Nex-N2.5 Pro vs mini: low prices meet the frozen planner test Pro completes 100 cases with 95 structured tool calls and 67% Sonnet alignment at a very low run cost. Mini's hosted route blocks tool requests, so its planner score remains unknown. Nex-N2.5-Pro costs about $0.15 for WebBrain's frozen 100-case planner test, but trails MiniMax M3 on Sonnet alignment. Mini's OpenRouter tool route rejects the test; Qwen 3.8 27B provides separate compact context. October 3, 2026 2026-10-03 Nex-N2.5-Pro Nex-N2.5-mini Nex AGI OpenRouter MiniMax M3 Qwen 3.8 27B browser agent benchmark tool calling">
<div class="post-meta">October 3, 2026 · 8 min read</div>
<div class="post-title">Nex-N2.5 Pro vs mini: low prices meet the frozen planner test</div>
<div class="post-excerpt">Pro completes 100 cases with 95 structured tool calls and 67% Sonnet alignment at a very low run cost. Mini's hosted route blocks tool requests, so its planner score remains unknown.</div>
</a>
<a href="/blog/spark-minicpm-compact-tool-routing" class="post-card" data-search="Spark X2.5 vs MiniCPM5: three compact browser-tool models tested Spark X2.5 vs MiniCPM5: three compact browser-tool models tested Three new official Q4 GGUF runs on the same saved Compact inputs, with strict action matching, loose tool-family scores, and the complete historical table. RTX 5090 tests of Spark-X2.5-4B, MiniCPM5-2B, and Spark-X2.5-1.7B, followed by a complete 15-model Compact comparison. September 26, 2026 2026-09-26 Spark-X2.5 MiniCPM5-2B browser agent benchmark tool calling RTX 5090">
<div class="post-meta">September 26, 2026 · 6 min read</div>
<div class="post-title">Spark X2.5 vs MiniCPM5: three compact browser-tool models tested</div>
<div class="post-excerpt">Three new official Q4 GGUF runs on the same saved Compact inputs, with strict action matching, loose tool-family scores, and the complete historical table.</div>
</a>
<a href="/blog/webbrain-36-8-jev-safesocial" class="post-card" data-search="Jev and SafeSocial arrive in WebBrain 36.8.0 Jev and SafeSocial arrive in WebBrain 36.8.0 WebBrain 36.8.0 adds two very different small-model features. Jev is a classifier-plus-LLM browser sidecar that did not produce a meaningful speedup in our tests. SafeSocial is our homegrown, public 13.6 MB local classifier that performs surprisingly well on moment-to-moment Instagram filtering. WebBrain 36.8.0 adds Jev for optional fast browser decisions and SafeSocial, our homegrown 13.6 MB local Instagram image classifier. Our results are deliberately mixed: Jev did not make our workloads faster, while the much smaller SafeSocial model works well for its narrow job. September 20, 2026 2026-09-20 WebBrain 36.8.0 Jev SafeSocial open model Hugging Face TypeSafe browser agent browser automation image classifier Instagram local AI EfficientNet-Lite0">
<div class="post-meta">September 20, 2026 · 8 min read</div>
<div class="post-title">Jev and SafeSocial arrive in WebBrain 36.8.0</div>
<div class="post-excerpt">WebBrain 36.8.0 adds two very different small-model features. Jev is a classifier-plus-LLM browser sidecar that did not produce a meaningful speedup in our tests. SafeSocial is our homegrown, public 13.6 MB local classifier that performs surprisingly well on moment-to-moment Instagram filtering.</div>
</a>
<a href="/blog/webbrain-compass-v2" class="post-card" data-search="WebBrain Compass Tiny v2: small model, sharper tool routing WebBrain Compass Tiny v2: small model, sharper tool routing Built on MiniCPM5-2B, Compass Tiny v2 reaches 16 exact actions and 41 tool-family matches out of 89 scenarios. A promising step toward WebBrain integration. Meet WebBrain Compass Tiny v2, our MiniCPM5-2B-based tool-routing model. Compare its merged BF16 results with 11 other models, including larger reference models. September 10, 2026 2026-09-10 WebBrain Compass Tiny v2 MiniCPM5-2B tool calling browser agent benchmark local AI">
<div class="post-meta">September 10, 2026 · 2 min read</div>
<div class="post-title">WebBrain Compass Tiny v2: small model, sharper tool routing</div>
<div class="post-excerpt">Built on MiniCPM5-2B, Compass Tiny v2 reaches 16 exact actions and 41 tool-family matches out of 89 scenarios. A promising step toward WebBrain integration.</div>
</a>
<a href="/blog/compact-tool-routing-models-compared" class="post-card" data-search="Compact tool routing compared: 10 models, from 2B to 35B Compact tool routing compared: 10 models, from 2B to 35B Six new local runs expand the Compact comparison. Qwen3.8-27B leads strict exact-action accuracy, while a looser name-only-inclusive metric produces a different ordering. An updated WebBrain Compact benchmark table comparing Compass Tiny with Gemma 4, Qwen 3.5, Qwen 3.6, Qwen 3.8, Nanbeige, and the LFM2.5 base model. September 7, 2026 2026-09-07 browser agent benchmark tool calling WebBrain Compact Gemma 4 Qwen 3.5 WebBrain Compass Tiny">
<div class="post-meta">September 7, 2026 · 3 min read</div>
<div class="post-title">Compact tool routing compared: 10 models, from 2B to 35B</div>
<div class="post-excerpt">Six new local runs expand the Compact comparison. Qwen3.8-27B leads strict exact-action accuracy, while a looser name-only-inclusive metric produces a different ordering.</div>
</a>
<a href="/blog/webbrain-compass-tiny-v1" class="post-card" data-search="Introducing WebBrain Compass Tiny: edge-sized Compact tool routing for WebBrain Introducing WebBrain Compass Tiny: edge-sized Compact tool routing for WebBrain A 2.6B model does not replace a 35B one at general reasoning. But inside WebBrain's narrow, production-shaped Compact routing loop, Compass Tiny v1 hits the same tool-name conformity as Qwen3.6-35B-A3B and beats its matched local base model by eight routing outcomes. WebBrain Compass Tiny v1 is a private preview of a 2.6B model built for WebBrain's Compact tool loop. In our first routing suite it matches Qwen3.6-35B-A3B on tool-name conformity and beats the local LFM2.5 base model it was trained from by eight routing outcomes, at a size meant for edge deployment. September 5, 2026 2026-09-05 WebBrain Compass Tiny browser agent tool calling compact mode edge AI LFM2.5 Qwen">
<div class="post-meta">September 5, 2026 · 6 min read</div>
<div class="post-title">Introducing WebBrain Compass Tiny: edge-sized Compact tool routing for WebBrain</div>
<div class="post-excerpt">A 2.6B model does not replace a 35B one at general reasoning. But inside WebBrain's narrow, production-shaped Compact routing loop, Compass Tiny v1 hits the same tool-name conformity as Qwen3.6-35B-A3B and beats its matched local base model by eight routing outcomes.</div>
</a>
<a href="/blog/deepseek-v4-flash-vision-openrouter" class="post-card" data-search="DeepSeek V4 Flash Vision Exp tested: a great planner with experimental eyes DeepSeek V4 Flash Vision Exp tested: a great planner with experimental eyes DeepSeek V4 Flash Vision Exp delivers the best first-action routing of our three new Flash models and a 2.60-second planner median. Its vision side is fast, but only 68 of 100 screenshots pass the strict rubric. We tested DeepSeek V4 Flash Vision Exp on 200 WebBrain browser-agent cases. Its text planning beats Qwen3.8 Flash and GLM-5.3 Flash, but its experimental vision route trails both. August 27, 2026 2026-08-27 DeepSeek V4 Flash Vision Exp DeepSeek V4 Flash 0731 OpenRouter multimodal model benchmark browser agent mixture of experts vision language model tool calling AI API pricing WebBrain">
<div class="post-meta">August 27, 2026 · 9 min read</div>
<div class="post-title">DeepSeek V4 Flash Vision Exp tested: a great planner with experimental eyes</div>
<div class="post-excerpt">DeepSeek V4 Flash Vision Exp delivers the best first-action routing of our three new Flash models and a 2.60-second planner median. Its vision side is fast, but only 68 of 100 screenshots pass the strict rubric.</div>
</a>
<a href="/blog/qwen38-flash-vs-glm53-flash" class="post-card" data-search="Qwen3.8 Flash vs GLM-5.3 Flash: the budget multimodal tier just reset Qwen3.8 Flash vs GLM-5.3 Flash: the budget multimodal tier just reset Qwen3.8 Flash reaches 79 strict vision passes and a 2.55-second planner median. GLM-5.3 Flash lands at 76 vision passes, follows the canonical first route more often, and costs half as much only while its 50% launch promotion lasts. We tested Qwen3.8 Flash and GLM-5.3 Flash on 400 WebBrain browser-agent cases. Both are huge multimodal MoEs with small active paths, million-token context, and near-identical list pricing. Qwen wins vision; GLM wins cost under a temporary launch discount. August 27, 2026 2026-08-27 Qwen3.8 Flash GLM-5.3 Flash OpenRouter multimodal model benchmark browser agent mixture of experts vision language model tool calling cheap AI API WebBrain">
<div class="post-meta">August 27, 2026 · 9 min read</div>
<div class="post-title">Qwen3.8 Flash vs GLM-5.3 Flash: the budget multimodal tier just reset</div>
<div class="post-excerpt">Qwen3.8 Flash reaches 79 strict vision passes and a 2.55-second planner median. GLM-5.3 Flash lands at 76 vision passes, follows the canonical first route more often, and costs half as much only while its 50% launch promotion lasts.</div>
</a>
<a href="/blog/muse-glimmer" class="post-card" data-search="Muse Glimmer Muse Glimmer Muse Glimmer reaches 73 strict vision passes and a 93.3% mean rubric score, ahead of the budget Qwen field. In text planning it emits 95 schema-valid tool calls, chooses 34 ideal tools, and sits above GPT-5.6 Luna on the common deterministic measures—but its 8.67-second median leaves a lot of speed on the table. Meta's open 30B Muse Glimmer beats every budget Qwen model in WebBrain's browser-vision benchmark, then lands above GPT-5.6 Luna on comparable first-action planner metrics—at the cost of much slower hosted inference. August 25, 2026 2026-08-25 Muse Glimmer Meta Muse Glimmer 30B OpenRouter vision model benchmark browser agent GPT-5.6 Luna Qwen3-VL tool calling open-weight AI local multimodal model WebBrain">
<div class="post-meta">August 25, 2026 · 12 min read</div>
<div class="post-title">Muse Glimmer</div>
<div class="post-excerpt">Muse Glimmer reaches 73 strict vision passes and a 93.3% mean rubric score, ahead of the budget Qwen field. In text planning it emits 95 schema-valid tool calls, chooses 34 ideal tools, and sits above GPT-5.6 Luna on the common deterministic measures—but its 8.67-second median leaves a lot of speed on the table.</div>
</a>
<a href="/blog/qwen-budget-vision-openrouter" class="post-card" data-search="Six budget Qwen vision models, compared: 600 requests, 54 cents on OpenRouter Six budget Qwen vision models, compared: 600 requests, 54 cents on OpenRouter Dense 32B Instruct takes the quality crown at 69 strict passes. The MoE 30B-A3B lands one pass behind at half the latency. Both thinking variants cost 4-5x more output tokens and didn't beat their instruct siblings — except at 8B, where thinking was worth 13 passes. We ran every current sub-dollar Qwen vision model on OpenRouter through WebBrain's 100-case browser-vision benchmark. Qwen3-VL-32B Instruct won on quality, the MoE 30B-A3B matched it at twice the speed, and the thinking variants billed four to five times the output tokens for little or no gain. August 22, 2026 2026-08-22 Qwen3-VL OpenRouter vision model benchmark browser agent qwen3.5-35b-a3b Qwen3-VL-32B Instruct Qwen3-VL-30B-A3B reasoning models vision language model WebBrain">
<div class="post-meta">August 22, 2026 · 7 min read</div>
<div class="post-title">Six budget Qwen vision models, compared: 600 requests, 54 cents on OpenRouter</div>
<div class="post-excerpt">Dense 32B Instruct takes the quality crown at 69 strict passes. The MoE 30B-A3B lands one pass behind at half the latency. Both thinking variants cost 4-5x more output tokens and didn't beat their instruct siblings — except at 8B, where thinking was worth 13 passes.</div>
</a>
<a href="/blog/tiny-vision-models-compared-ii" class="post-card" data-search="Tiny Vision Models Compared II: WebBrain, Gemma 4, and Larger Qwen 3.5 Models Join the Race Tiny Vision Models Compared II: WebBrain, Gemma 4, and Larger Qwen 3.5 Models Join the Race A browser-specific 450M fine-tune reaches the same strict-pass band as Qwen 3.5 2B and Gemma 4 E4B while remaining small enough for WebGPU. It will become WebBrain's default local vision model. WebBrain VL 2 450M joins our 100-case browser-vision benchmark alongside Qwen 3.5 2B and 4B and Gemma 4 E4B, with the earlier Qwen and LFM baselines retained for context. August 22, 2026 2026-08-22 WebBrain VL 2 450M WebBrain Qwen 3.5 4B Qwen 3.5 2B Qwen 3.5 0.8B Gemma 4 E4B LFM2.5-VL WebGPU browser vision vision-language model">
<div class="post-meta">August 22, 2026 · 8 min read</div>
<div class="post-title">Tiny Vision Models Compared II: WebBrain, Gemma 4, and Larger Qwen 3.5 Models Join the Race</div>
<div class="post-excerpt">A browser-specific 450M fine-tune reaches the same strict-pass band as Qwen 3.5 2B and Gemma 4 E4B while remaining small enough for WebGPU. It will become WebBrain's default local vision model.</div>
</a>
<a href="/blog/fine-tuning-lfm25-vl-450m-for-browser-tasks" class="post-card" data-search="From 16K to 50K Browser Screenshots: Fine-Tuning LFM2.5-VL-450M From 16K to 50K Browser Screenshots: Fine-Tuning LFM2.5-VL-450M The raw 450M model started near zero on WebBrain's strict browser contract. Two data rounds moved it to 44/100 in the deployed WebGPU package—above a 3B LFM run and into the same band as much larger Qwen and Gemma models. How WebBrain turned Liquid AI's 450M vision-language model into a browser specialist using a reviewed 16,646-example first dataset, a targeted 50,000-example second dataset, LoRA, and a held-out 100-case test. August 22, 2026 2026-08-22 WebBrain VL 2 450M LFM2.5-VL-450M Liquid AI fine-tuning LoRA browser screenshots GUI dataset WebGPU ONNX teacher distillation">
<div class="post-meta">August 22, 2026 · 10 min read</div>
<div class="post-title">From 16K to 50K Browser Screenshots: Fine-Tuning LFM2.5-VL-450M</div>
<div class="post-excerpt">The raw 450M model started near zero on WebBrain's strict browser contract. Two data rounds moved it to 44/100 in the deployed WebGPU package—above a 3B LFM run and into the same band as much larger Qwen and Gemma models.</div>
</a>
<a href="/blog/tiny-vision-models-qwen38-reference" class="post-card" data-search="Tiny Vision Models Compared: LFM 2.5 VL vs Qwen 3.5 0.8B Tiny Vision Models Compared: LFM 2.5 VL vs Qwen 3.5 0.8B Qwen 3.8 27B is clearly better. The surprise is that Qwen 3.5 0.8B and tiny LFM2.5-VL models can still recover useful browser state, especially when vision is a local fallback rather than the whole agent. Qwen 3.8 27B set a strong 74/100 reference in WebBrain's 100-case browser-vision benchmark. We compare it with Qwen 3.5 0.8B and Liquid AI's 450M and 1.6B local VLMs—and explain why a weak benchmark score can still describe a useful fallback. August 21, 2026 2026-08-21 WebBrain Qwen 3.8 27B Qwen 3.5 0.8B LFM2.5-VL-450M LFM2.5-VL-1.6B local vision model WebGPU browser agent Apocalypse Mode vision benchmark">
<div class="post-meta">August 21, 2026 · 7 min read</div>
<div class="post-title">Tiny Vision Models Compared: LFM 2.5 VL vs Qwen 3.5 0.8B</div>
<div class="post-excerpt">Qwen 3.8 27B is clearly better. The surprise is that Qwen 3.5 0.8B and tiny LFM2.5-VL models can still recover useful browser state, especially when vision is a local fallback rather than the whole agent.</div>
</a>
<a href="/blog/why-webbrain-33-is-gpl" class="post-card" data-search="Why WebBrain 33.0.0 is GPL Why WebBrain 33.0.0 is GPL 33.0.0 is GPL because full-text Wikipedia search in the extension needs a GPL runtime in the package. Earlier tags stay MIT. WebBrain 33.0.0 ships under GPL-3.0-or-later because the extension bundles the Xapian/libzim WebAssembly runtime for full-text Wikipedia ZIM search. Releases before 33.0.0 remain MIT. August 20, 2026 2026-08-20 WebBrain GPL GPL-3.0 open source license Xapian libzim Wikipedia offline Apocalypse Mode">
<div class="post-meta">August 20, 2026 · 2 min read</div>
<div class="post-title">Why WebBrain 33.0.0 is GPL</div>
<div class="post-excerpt">33.0.0 is GPL because full-text Wikipedia search in the extension needs a GPL runtime in the package. Earlier tags stay MIT.</div>
</a>
<a href="/blog/why-we-built-apocalypse-mode" class="post-card" data-search="The world needs offline access. That's why we built Apocalypse Mode. The world needs offline access. That's why we built Apocalypse Mode. The internet feels permanent until it isn't. Apocalypse Mode prepares a self-contained local core—WebGPU text and vision models, offline Wikipedia, and an Emergency Box of practical documents—before connectivity disappears. Disasters, wars, infrastructure failures, and the unexpected can take the internet away. WebBrain's Apocalypse Mode keeps local AI, Wikipedia, and practical references ready in the browser. August 17, 2026 2026-08-17 Apocalypse Mode offline access offline AI WebGPU disaster preparedness Wikipedia offline local LLM Emergency Box WebBrain">
<div class="post-meta">August 17, 2026 · 6 min read</div>
<div class="post-title">The world needs offline access. That's why we built Apocalypse Mode.</div>
<div class="post-excerpt">The internet feels permanent until it isn't. Apocalypse Mode prepares a self-contained local core—WebGPU text and vision models, offline Wikipedia, and an Emergency Box of practical documents—before connectivity disappears.</div>
</a>
<a href="/blog/webbrain-31-offline-webgpu-vision" class="post-card" data-search="WebBrain 31 brings offline WebGPU vision to text-only models WebBrain 31 brings offline WebGPU vision to text-only models WebBrain 31.0.0 can run a lightweight vision model locally through WebGPU, then pass its visual observations to the original text-only planner. That makes inexpensive models such as DeepSeek V4 Flash and Laguna S 2.1 more capable without turning them into larger multimodal checkpoints. WebBrain 31.0.0 adds optional offline vision powered by WebGPU and Liquid AI's sub-1GB LFM2.5-VL-450M, giving strong text-only models a local visual layer. August 14, 2026 2026-08-14 WebBrain 31 WebGPU offline vision Liquid AI LFM2.5-VL-450M DeepSeek V4 Flash Laguna S 2.1 local AI browser agent text-only model">
<div class="post-meta">August 14, 2026 · 5 min read</div>
<div class="post-title">WebBrain 31 brings offline WebGPU vision to text-only models</div>
<div class="post-excerpt">WebBrain 31.0.0 can run a lightweight vision model locally through WebGPU, then pass its visual observations to the original text-only planner. That makes inexpensive models such as DeepSeek V4 Flash and Laguna S 2.1 more capable without turning them into larger multimodal checkpoints.</div>
</a>
<a href="/blog/exl3-sparkinfer-blackwell-lite" class="post-card" data-search="EXL3 is the best quantization format for local AI right now — especially on Blackwell lite EXL3 is the best quantization schema for local AI right now EXL3 keeps more quality per bit than GGUF or EXL2, converts cheaply, and its trellis GEMMs pair perfectly with Blackwell sm_120/sm_121 hardware. SparkInfer (vLLM fork) integrated it natively and we're getting ~3x the throughput of ExLlamaV3 — with sane tensor/expert parallelism even on odd card counts. EXL3 (turboderp's QTIP-derived trellis format) is the best low-bit quantization for local AI on Blackwell lite hardware. Now it runs ~3x faster through SparkInfer, the vLLM fork — including on odd GPU counts and DGX Spark. August 11, 2026 2026-08-11 EXL3 quantization SparkInfer vLLM ExLlamaV3 Blackwell DGX Spark RTX 5090 trellis QTIP local LLM MoE tensor parallel">
<div class="post-meta">August 11, 2026 · 6 min read</div>
<div class="post-title">EXL3 is the best quantization schema for local AI right now</div>
<div class="post-excerpt">EXL3 keeps more quality per bit than GGUF or EXL2, converts cheaply, and its trellis GEMMs pair perfectly with Blackwell sm_120/sm_121 hardware. SparkInfer (vLLM fork) integrated it natively and we're getting ~3x the throughput of ExLlamaV3 — with sane tensor/expert parallelism even on odd card counts.</div>
</a>
<a href="/blog/mcp-server-introduction" class="post-card" data-search="WebBrain MCP: Give Claude Code, OpenCode, and Codex access to your real browser session WebBrain MCP: Give Claude Code, OpenCode, and Codex access to your real browser session The WebBrain MCP server bridges Claude Code, OpenCode, and Codex to your real, already-signed-in browser. No headless login walls. No cookie transfers. Just ask your coding agent to interact with any page you're logged into. The WebBrain MCP server lets any MCP-capable coding agent — Claude Code, OpenCode, Codex, Cursor — delegate browser tasks to your real, already-authenticated browser. No headless login dances, no cookie transfers. Just ask your agent to fill out a form, check a dashboard, or read a page behind SSO, and it runs in the browser you actually use. August 8, 2026 2026-08-08 WebBrain MCP Model Context Protocol Claude Code OpenCode Codex Cursor browser agent automation authentication headless browser local LLM Chrome extension">
<div class="post-meta">August 8, 2026 · 11 min read</div>
<div class="post-title">WebBrain MCP: Give Claude Code, OpenCode, and Codex access to your real browser session</div>
<div class="post-excerpt">The WebBrain MCP server bridges Claude Code, OpenCode, and Codex to your real, already-signed-in browser. No headless login walls. No cookie transfers. Just ask your coding agent to interact with any page you're logged into.</div>
</a>
<a href="/blog/qwen36-27b-vs-laguna-xs-21-vision" class="post-card" data-search="Qwen 3.6 27B vs Laguna XS 2.1 Vision: which local LLM is better? Qwen 3.6 27B vs Laguna XS 2.1 Vision: which local LLM is better? Qwen remains the safer ready-to-run local multimodal model, but Laguna XS 2.1's new MoonViT vision bridge turns Poolside's 3B-active MoE into a serious contender—and brings American and Chinese open-weight models much closer to head-to-head. Qwen 3.6 27B and Laguna XS 2.1 take radically different routes to local multimodal AI. We compare their architectures, WebBrain planner results, hardware tradeoffs, and Laguna's new experimental vision checkpoint. August 5, 2026 2026-08-05 WebBrain Qwen 3.6 27B Laguna XS 2.1 Laguna XS 2.1 Vision Poolside local LLM local VLM NVFP4 MoonViT mixture of experts open-weight AI browser agent">
<div class="post-meta">August 5, 2026 · 11 min read</div>
<div class="post-title">Qwen 3.6 27B vs Laguna XS 2.1 Vision: which local LLM is better?</div>
<div class="post-excerpt">Qwen remains the safer ready-to-run local multimodal model, but Laguna XS 2.1's new MoonViT vision bridge turns Poolside's 3B-active MoE into a serious contender—and brings American and Chinese open-weight models much closer to head-to-head.</div>
</a>
<a href="/blog/deepseek-v4-flash-vision-support" class="post-card" data-search="DeepSeek V4 Flash now has vision support DeepSeek V4 Flash now has vision support DeepSeek V4 Flash can now inspect screenshots, interfaces, layouts, and other visual browser context—while retaining the strong price-performance profile we saw in our internal tests. WebBrain's DeepSeek V4 Flash Vision NVFP4 release adds screenshot and interface understanding to a model that was previously text-only. August 4, 2026 2026-08-04 WebBrain DeepSeek V4 Flash DeepSeek V4 Flash Vision NVFP4 vision-language model multimodal model browser vision browser agent open-weight AI">
<div class="post-meta">August 4, 2026 · 2 min read</div>
<div class="post-title">DeepSeek V4 Flash now has vision support</div>
<div class="post-excerpt">DeepSeek V4 Flash can now inspect screenshots, interfaces, layouts, and other visual browser context—while retaining the strong price-performance profile we saw in our internal tests.</div>
</a>
<a href="/blog/inkling-small-glm-5-2-vision-benchmark" class="post-card" data-search="Two ways to give an open model eyes: Inkling Small vs GLM-5.2 Vision Two ways to give an open model eyes: Inkling Small vs GLM-5.2 Vision Both models read all 12 target strings and found the real sign-in blocker. Inkling Small was faster and more accurate on focus and error-state cues; the self-hosted GLM-5.2 derivative proved that a third-party vision graft can make a text-first 744B-class backbone genuinely useful on screenshots. We ran the same frozen WebBrain UI-vision probe against Thinking Machines Inkling Small on OpenRouter and a full GLM-5.2 Vision hybrid on four RTX PRO 6000 Blackwell GPUs at RunPod. August 2, 2026 2026-08-02 WebBrain Inkling Small Thinking Machines GLM-5.2 Vision Baseten RunPod OpenRouter multimodal model vision-language model open-weight AI browser agent UI understanding">
<div class="post-meta">August 2, 2026 · 10 min read</div>
<div class="post-title">Two ways to give an open model eyes: Inkling Small vs GLM-5.2 Vision</div>
<div class="post-excerpt">Both models read all 12 target strings and found the real sign-in blocker. Inkling Small was faster and more accurate on focus and error-state cues; the self-hosted GLM-5.2 derivative proved that a third-party vision graft can make a text-first 744B-class backbone genuinely useful on screenshots.</div>
</a>
<a href="/blog/american-chinese-open-model-frontier-gap-benchmark" class="post-card" data-search="Two AI gaps are narrowing: American open weights and Chinese frontier models Two AI gaps are narrowing: American open weights and Chinese frontier models DeepSeek V4 Flash and Tencent HY3 lead our reference-free first-action consensus, while Gemini 3.6 Flash goes 100-for-100 on valid calls and Poolside Laguna XS becomes a remarkably cheap American open-weight contender. GLM-5.2's released Vision NVFP4 graft changes its modality story, and Qwen 3.6 27B remains our practical RTX 5090 pick. A consensus-ranked, cost-aware WebBrain planner test of thirteen OpenRouter models from OpenAI, Anthropic, Google, xAI, Poolside, MiniMax, Z.ai, DeepSeek, Thinking Machines, Qwen, Moonshot, and Tencent. August 2, 2026 2026-08-02 WebBrain OpenRouter open-weight AI American AI Chinese AI DeepSeek V4 Tencent HY3 GLM-5.2 Kimi K3 MiniMax M3 Claude Sonnet 5 GPT-5.6 Qwen 3.6 27B Gemini 3.6 Flash Grok 4.5 Poolside Laguna XS browser agent tool calling">
<div class="post-meta">August 2, 2026 · 22 min read</div>
<div class="post-title">Two AI gaps are narrowing: American open weights and Chinese frontier models</div>
<div class="post-excerpt">DeepSeek V4 Flash and Tencent HY3 lead our reference-free first-action consensus, while Gemini 3.6 Flash goes 100-for-100 on valid calls and Poolside Laguna XS becomes a remarkably cheap American open-weight contender. GLM-5.2's released Vision NVFP4 graft changes its modality story, and Qwen 3.6 27B remains our practical RTX 5090 pick.</div>
</a>
<a href="/blog/thinking-machines-inkling-openrouter-planner-benchmark" class="post-card" data-search="Inkling makes the American open-weight comeback genuinely multimodal Inkling makes the American open-weight comeback genuinely multimodal Inkling produced 96 valid calls and chose the ideal tool 45 times, while reaching 73% Sonnet alignment. Maximum reasoning did not help—but working image and audio input make this 975B American open-weight release unusually complete. We tested Thinking Machines Inkling through OpenRouter on WebBrain's frozen 100-case browser-planner benchmark, then checked its image and audio understanding. July 23, 2026 2026-07-23 WebBrain Thinking Machines Inkling OpenRouter open-weight model American AI multimodal AI audio model browser agent tool calling Poolside Laguna">
<div class="post-meta">July 23, 2026 · 10 min read</div>
<div class="post-title">Inkling makes the American open-weight comeback genuinely multimodal</div>
<div class="post-excerpt">Inkling produced 96 valid calls and chose the ideal tool 45 times, while reaching 73% Sonnet alignment. Maximum reasoning did not help—but working image and audio input make this 975B American open-weight release unusually complete.</div>
</a>
<a href="/blog/nanbeige-42-3b-bf16-planner-benchmark" class="post-card" data-search="Nanbeige 4.2 3B punches above its size, but Qwen 3.5 9B still wins our frozen planner test Nanbeige 4.2 3B punches above its size, but Qwen 3.5 9B still wins our frozen planner test Nanbeige 4.2 3B produced 90 parsed calls, reached 67% Sonnet alignment, and delivered a 1.49s single-request median. That is impressive for 4B total parameters, but below Qwen 3.5 9B and Gemma 4 E4B in our first-action test. We tested Nanbeige4.2-3B in BF16 on an RTX 5090. Its official agent benchmarks beat Qwen 3.5 9B and Gemma 4 12B, but WebBrain's frozen browser-planner replay did not reproduce that lead. July 22, 2026 2026-07-22 WebBrain Nanbeige Nanbeige 4.2 3B Qwen 3.5 9B Gemma 4 12B Gemma 4 E4B vLLM BF16 local LLM browser agent tool calling">
<div class="post-meta">July 22, 2026 · 8 min read</div>
<div class="post-title">Nanbeige 4.2 3B punches above its size, but Qwen 3.5 9B still wins our frozen planner test</div>
<div class="post-excerpt">Nanbeige 4.2 3B produced 90 parsed calls, reached 67% Sonnet alignment, and delivered a 1.49s single-request median. That is impressive for 4B total parameters, but below Qwen 3.5 9B and Gemma 4 E4B in our first-action test.</div>
</a>
<a href="/blog/poolside-laguna-m1-openrouter-planner-benchmark" class="post-card" data-search="Poolside Laguna M.1 reaches 73% in WebBrain, but 225B does not win Poolside Laguna M.1 reaches 73% in WebBrain, but 225B does not win Laguna M.1 returned 92 valid tool calls and reached 73% Sonnet alignment with a high-reasoning request. The larger model is much more dispatchable than Laguna S, but exact and ideal tool selection barely improve. Poolside Laguna M.1 is almost twice the size of Laguna S 2.1 and fixes much of its no-tool problem. It reached 73% Sonnet alignment in WebBrain, tying Hy3's headline score but still trailing MiniMax M3 and the strongest 27–31B local planners. July 21, 2026 2026-07-21 WebBrain Poolside Laguna M.1 OpenRouter Laguna S 2.1 Tencent Hy3 MiniMax M3 Gemma 4 ThinkingCap open-weight model browser agent planner benchmark tool calling">
<div class="post-meta">July 21, 2026 · 9 min read</div>
<div class="post-title">Poolside Laguna M.1 reaches 73% in WebBrain, but 225B does not win</div>
<div class="post-excerpt">Laguna M.1 returned 92 valid tool calls and reached 73% Sonnet alignment with a high-reasoning request. The larger model is much more dispatchable than Laguna S, but exact and ideal tool selection barely improve.</div>
</a>
<a href="/blog/poolside-laguna-s-openrouter-planner-benchmark" class="post-card" data-search="Poolside Laguna S 2.1 reaches 71% with a high-reasoning request, but still trails Hy3 Poolside Laguna S 2.1 reaches 71% with a high-reasoning request, but still trails Hy3 Laguna S 2.1 is much cleaner than Laguna XS and extraordinarily cheap on OpenRouter. A high-reasoning request lifted Sonnet alignment from 65% to 71%, but reduced parsed tool calls from 86 to 78 and still did not catch Hy3 or MiniMax M3. Poolside's Laguna S 2.1 beats Hy3 on two published SWE benchmarks and MiniMax M3 on Terminal-Bench 2.1. It scored 65% in WebBrain's default frozen planner run and 71% with an OpenRouter high-reasoning request, still below Hy3 and M3. July 21, 2026 2026-07-21 WebBrain Poolside Laguna S 2.1 OpenRouter Tencent Hy3 MiniMax M3 Inkling open-weight model browser agent planner benchmark tool calling">
<div class="post-meta">July 21, 2026 · 9 min read</div>
<div class="post-title">Poolside Laguna S 2.1 reaches 71% with a high-reasoning request, but still trails Hy3</div>
<div class="post-excerpt">Laguna S 2.1 is much cleaner than Laguna XS and extraordinarily cheap on OpenRouter. A high-reasoning request lifted Sonnet alignment from 65% to 71%, but reduced parsed tool calls from 86 to 78 and still did not catch Hy3 or MiniMax M3.</div>
</a>
<a href="/blog/glm52-nvidia-webbrain-planner-benchmark" class="post-card" data-search="GLM-5.2 is not WebBrain's new planner reference yet GLM-5.2 is not WebBrain's new planner reference yet GLM-5.2 completed four 100-case WebBrain planner suites with zero transport errors after resumed runs. The frozen row lands at 69% Sonnet alignment, 21 exact first calls, and 36 ideal tool-name matches. We ran z-ai/glm-5.2 through WebBrain's frozen, full, mid, and compact browser-agent first-tool planner suites on NVIDIA. It should be a stronger frontier open-model candidate than the old Sonnet reference, but this narrow planner result is more cautious. July 14, 2026 2026-07-14 WebBrain GLM-5.2 z-ai NVIDIA browser agent planner benchmark tool calling frontier model">
<div class="post-meta">July 14, 2026 · 6 min read</div>
<div class="post-title">GLM-5.2 is not WebBrain's new planner reference yet</div>
<div class="post-excerpt">GLM-5.2 completed four 100-case WebBrain planner suites with zero transport errors after resumed runs. The frozen row lands at 69% Sonnet alignment, 21 exact first calls, and 36 ideal tool-name matches.</div>
</a>
<a href="/blog/webmcp-integration" class="post-card" data-search="WebMCP: Websites as Tools for AI Agents, and Why We're Excited WebMCP: Websites as Tools for AI Agents, and Why We're Excited A new browser standard lets websites hand AI agents a menu of actions instead of forcing them to squint at screenshots. One GitHub issue sparked a conversation that could change how WebBrain works with the web. WebMCP lets websites expose structured tools to AI agents. A conversation with Google's Gemini-in-Chrome team is pushing us toward integration, and we couldn't be more excited. July 14, 2026 2026-07-14 WebMCP browser agent AI agent tools WebBrain Model Context Protocol agentic web W3C standard web automation">
<div class="post-meta">July 14, 2026 · 5 min read</div>
<div class="post-title">WebMCP: Websites as Tools for AI Agents, and Why We're Excited</div>
<div class="post-excerpt">A new browser standard lets websites hand AI agents a menu of actions instead of forcing them to squint at screenshots. One GitHub issue sparked a conversation that could change how WebBrain works with the web.</div>
</a>
<a href="/blog/poolside-laguna-xs-openrouter-planner-benchmark" class="post-card" data-search="Poolside Laguna XS is fast and free, but not a WebBrain planner win Poolside Laguna XS is fast and free, but not a WebBrain planner win Laguna XS completed WebBrain's frozen planner suite with 92 parsed calls, 89 valid tool names, 65% Sonnet alignment, and a 1.30s median latency. The endpoint is quick; the planner quality is not competitive. We ran poolside/laguna-xs-2.1:free through WebBrain's frozen 100-case browser-agent first-tool benchmark on OpenRouter. It is very fast and free to call, but 65% Sonnet alignment, only 6 exact first actions, and malformed tool names keep it out of the planner shortlist. July 10, 2026 2026-07-10 WebBrain Poolside Laguna XS OpenRouter browser agent planner benchmark tool calling free model">
<div class="post-meta">July 10, 2026 · 6 min read</div>
<div class="post-title">Poolside Laguna XS is fast and free, but not a WebBrain planner win</div>
<div class="post-excerpt">Laguna XS completed WebBrain's frozen planner suite with 92 parsed calls, 89 valid tool names, 65% Sonnet alignment, and a 1.30s median latency. The endpoint is quick; the planner quality is not competitive.</div>
</a>
<a href="/blog/thinkingcap-qwen36-27b-planner-benchmark" class="post-card" data-search="ThinkingCap Qwen 3.6 27B is a serious WebBrain planner candidate ThinkingCap Qwen 3.6 27B is a serious WebBrain planner candidate ThinkingCap lands at 77% Sonnet alignment, 91 parsed tool calls, 19 exact first actions, and 2.25s median latency. It looks like a real local planner candidate, but not a clean replacement for the Qwen baselines. We ran ThinkingCap-Qwen3.6-27B INT4 AutoRound through WebBrain's frozen browser-agent first-tool benchmark. The model tied Qwen 3.6 27B on all-case Sonnet alignment, returned in 2.25s median latency, and exposed a sharper boundary-behavior tradeoff. July 9, 2026 2026-07-09 WebBrain ThinkingCap Qwen 3.6 Qwen 3.6 27B BottleCap AI vLLM browser agent planner benchmark tool calling">
<div class="post-meta">July 9, 2026 · 8 min read</div>
<div class="post-title">ThinkingCap Qwen 3.6 27B is a serious WebBrain planner candidate</div>
<div class="post-excerpt">ThinkingCap lands at 77% Sonnet alignment, 91 parsed tool calls, 19 exact first actions, and 2.25s median latency. It looks like a real local planner candidate, but not a clean replacement for the Qwen baselines.</div>
</a>
<a href="/blog/nex-n2-mini-openrouter-planner-benchmark" class="post-card" data-search="Nex-N2-mini is cheap and fast, but not a WebBrain planner win Nex-N2-mini is cheap and fast, but not a WebBrain planner win Nex-N2-mini completed WebBrain's frozen planner suite with 93 parsed tool calls, a 2.23s median latency, and about $0.045 reported OpenRouter cost, but only 65% Sonnet alignment. We ran nex-agi/nex-n2-mini through WebBrain's frozen 100-case browser-agent first-tool benchmark on OpenRouter. The 35B MoE endpoint is fast and extremely cheap, but its planner quality lands well below the current shortlist. July 8, 2026 2026-07-08 WebBrain Nex-N2-mini Nex AGI OpenRouter browser agent planner benchmark tool calling Qwen 3.7 Plus Agents-A1 Tencent Hy3">
<div class="post-meta">July 8, 2026 · 6 min read</div>
<div class="post-title">Nex-N2-mini is cheap and fast, but not a WebBrain planner win</div>
<div class="post-excerpt">Nex-N2-mini completed WebBrain's frozen planner suite with 93 parsed tool calls, a 2.23s median latency, and about $0.045 reported OpenRouter cost, but only 65% Sonnet alignment.</div>
</a>
<a href="/blog/nemotron3-ultra-openrouter-planner-benchmark" class="post-card" data-search="Nemotron 3 Ultra is huge, free, and not a WebBrain planner win Nemotron 3 Ultra is huge, free, and not a WebBrain planner win Nemotron 3 Ultra completed WebBrain's frozen planner benchmark through OpenRouter's free endpoint, but 81 parsed calls, 65% Sonnet alignment, and a 40.6s p95 keep it out of the planner shortlist. We ran nvidia/nemotron-3-ultra-550b-a55b:free through WebBrain's frozen 100-case browser-agent first-tool benchmark on OpenRouter. The 550B-total MoE model completed the run, but landed well below the best planner rows. July 8, 2026 2026-07-08 WebBrain Nemotron 3 Ultra NVIDIA OpenRouter browser agent planner benchmark tool calling Qwen 3.7 Plus Agents-A1 Tencent Hy3">
<div class="post-meta">July 8, 2026 · 6 min read</div>
<div class="post-title">Nemotron 3 Ultra is huge, free, and not a WebBrain planner win</div>
<div class="post-excerpt">Nemotron 3 Ultra completed WebBrain's frozen planner benchmark through OpenRouter's free endpoint, but 81 parsed calls, 65% Sonnet alignment, and a 40.6s p95 keep it out of the planner shortlist.</div>
</a>
<a href="/blog/agents-a1-webbrain-planner-benchmark" class="post-card" data-search="Agents-A1 in WebBrain's frozen planner benchmark Agents-A1 in WebBrain's frozen planner benchmark Agents-A1 beats Qwen 3.6 35B-A3B on Sonnet alignment and local latency in WebBrain's frozen planner harness, but it does not beat Qwen on exact or ideal first-call matching. We ran Agents-A1 AWQ INT4 through WebBrain's frozen 100-case browser-agent first-tool benchmark to test whether its agentic benchmark claims transfer to WebBrain's planner workload. July 8, 2026 2026-07-08 WebBrain Agents-A1 InternScience Qwen 3.6 Ornith Gemma 4 browser agent planner benchmark tool calling">
<div class="post-meta">July 8, 2026 · 7 min read</div>
<div class="post-title">Agents-A1 in WebBrain's frozen planner benchmark</div>
<div class="post-excerpt">Agents-A1 beats Qwen 3.6 35B-A3B on Sonnet alignment and local latency in WebBrain's frozen planner harness, but it does not beat Qwen on exact or ideal first-call matching.</div>
</a>
<a href="/blog/five-things-you-shouldnt-do-with-webbrain" class="post-card" data-search="Five things you shouldn't do with WebBrain Five things you shouldn't do with WebBrain We briefly interrupt the benchmark series for a video-sized warning label: WebBrain can automate a browser, but consent, legality, and basic human decency still apply. A short pause from WebBrain's LLM benchmark posts: five things a browser agent can technically help with, but absolutely should not be used for. July 7, 2026 2026-07-07 WebBrain browser agent AI automation AI safety responsible automation YouTube">
<div class="post-meta">July 7, 2026 · 3 min read</div>
<div class="post-title">Five things you shouldn't do with WebBrain</div>
<div class="post-excerpt">We briefly interrupt the benchmark series for a video-sized warning label: WebBrain can automate a browser, but consent, legality, and basic human decency still apply.</div>
</a>
<a href="/blog/tencent-hy3-openrouter-planner-benchmark" class="post-card" data-search="Tencent Hy3 is a great OpenRouter planner, but text-only for now Tencent Hy3 is a great OpenRouter planner, but text-only for now Tencent Hy3 reached 95/100 parsed tool calls, 20/100 exact first-action matches, and 73% Sonnet alignment. It is a strong OpenRouter planner row, with multimodality as the missing piece. We ran Tencent Hy3 through WebBrain's frozen 100-case browser-agent planner benchmark on OpenRouter. Hy3 lands in the serious hosted planner tier with 95 parsed tool calls, 20 exact first-action matches, and 73% Sonnet alignment, but the current model is still text-only. July 7, 2026 2026-07-07 WebBrain Tencent Hy3 OpenRouter MiniMax M3 MiniMax M2.7 Qwen 3.7 Plus browser agent planner benchmark tool calling">
<div class="post-meta">July 7, 2026 · 6 min read</div>
<div class="post-title">Tencent Hy3 is a great OpenRouter planner, but text-only for now</div>
<div class="post-excerpt">Tencent Hy3 reached 95/100 parsed tool calls, 20/100 exact first-action matches, and 73% Sonnet alignment. It is a strong OpenRouter planner row, with multimodality as the missing piece.</div>
</a>
<a href="/blog/ollama-launch-handoff" class="post-card" data-search="WebBrain now has an Ollama launch handoff WebBrain now has an Ollama launch handoff WebBrain can now be configured from an Ollama launch command: choose a local model, open the WebBrain handoff page, confirm the browser prompt, and the extension switches to Ollama automatically. It is not in upstream Ollama yet, but you can try it from the branch today. A new Ollama launch handoff can configure WebBrain for a local Ollama model from one terminal command. The integration is not upstream in Ollama yet, but it is available today from the esokullu/ollama branch while we work toward official integration. July 2, 2026 2026-07-02 WebBrain Ollama local LLM browser agent open source Chrome extension Firefox extension Ollama launch">
<div class="post-meta">July 2, 2026 · 4 min read</div>
<div class="post-title">WebBrain now has an Ollama launch handoff</div>
<div class="post-excerpt">WebBrain can now be configured from an Ollama launch command: choose a local model, open the WebBrain handoff page, confirm the browser prompt, and the extension switches to Ollama automatically. It is not in upstream Ollama yet, but you can try it from the branch today.</div>
</a>
<a href="/blog/qwen37-plus-openrouter-planner-benchmark" class="post-card" data-search="Qwen 3.7 Plus is a serious OpenRouter rival to MiniMax M3 Qwen 3.7 Plus is a serious OpenRouter rival to MiniMax M3 Qwen 3.7 Plus reached 75/100 Sonnet alignment, 95/100 parsed native tool calls, and the best ideal tool-name score in the top hosted slice of WebBrain's frozen planner table. We ran qwen/qwen3.7-plus through WebBrain's frozen 100-case browser-agent planner benchmark on OpenRouter. It ties MiniMax M3 on all-case Sonnet alignment, beats it on tool-call discipline, and lands just below the Gemma 4 31B QAT, Qwen 3.6 27B, and MiniMax M2.7 cluster. July 2, 2026 2026-07-02 WebBrain Qwen 3.7 Plus OpenRouter MiniMax M3 MiniMax M2.7 Gemma 4 31B Qwen 3.6 27B browser agent planner benchmark tool calling">
<div class="post-meta">July 2, 2026 · 6 min read</div>
<div class="post-title">Qwen 3.7 Plus is a serious OpenRouter rival to MiniMax M3</div>
<div class="post-excerpt">Qwen 3.7 Plus reached 75/100 Sonnet alignment, 95/100 parsed native tool calls, and the best ideal tool-name score in the top hosted slice of WebBrain's frozen planner table.</div>
</a>
<a href="/blog/qwen36-27b-nvfp4-planner-benchmark" class="post-card" data-search="Qwen 3.6 27B NVFP4 is much faster, but not a clean planner upgrade Qwen 3.6 27B NVFP4 is much faster, but not a clean planner upgrade Qwen 3.6 27B NVFP4 gives WebBrain a big local latency win: 96/100 parsed calls, 1.8s median latency, and a top-10 Sonnet-reference result. The planner quality story is more mixed. We tested nvidia/Qwen3.6-27B-NVFP4 against WebBrain's frozen browser-agent first-tool benchmark. It is far faster than the older Qwen 3.6 27B row and enters the top 10, but loses three all-case Sonnet-alignment points on boundary cases. July 2, 2026 2026-07-02 WebBrain Qwen 3.6 Qwen 3.6 27B NVFP4 NVIDIA vLLM browser agent planner benchmark tool calling">
<div class="post-meta">July 2, 2026 · 6 min read</div>
<div class="post-title">Qwen 3.6 27B NVFP4 is much faster, but not a clean planner upgrade</div>
<div class="post-excerpt">Qwen 3.6 27B NVFP4 gives WebBrain a big local latency win: 96/100 parsed calls, 1.8s median latency, and a top-10 Sonnet-reference result. The planner quality story is more mixed.</div>
</a>
<a href="/blog/ornith-35b-webbrain-planner-benchmark" class="post-card" data-search="Ornith-1.0-35B in WebBrain's frozen planner benchmark Ornith-1.0-35B in WebBrain's frozen planner benchmark Ornith-1.0-35B is a strong local planner and narrowly beats Qwen 3.6 35B on WebBrain's all-case Sonnet alignment. It does not beat Gemma 4 31B QAT in this browser-agent harness. We ran Ornith-1.0-35B through WebBrain's frozen 100-case browser-agent first-tool benchmark after DeepReinforce's model card claimed wins over Qwen3.6-35B and Gemma4-31B on agentic coding benchmarks. June 26, 2026 2026-06-26 WebBrain Ornith Ornith-1.0-35B DeepReinforce browser agent tool calling planner benchmark Gemma 4 Qwen 3.6">
<div class="post-meta">June 26, 2026 · 6 min read</div>
<div class="post-title">Ornith-1.0-35B in WebBrain's frozen planner benchmark</div>
<div class="post-excerpt">Ornith-1.0-35B is a strong local planner and narrowly beats Qwen 3.6 35B on WebBrain's all-case Sonnet alignment. It does not beat Gemma 4 31B QAT in this browser-agent harness.</div>
</a>
<a href="/blog/tiny-lfm25-raw-planner-benchmark" class="post-card" data-search="Tiny raw LFM 2.5 checkpoints in WebBrain's frozen planner benchmark Tiny raw LFM 2.5 checkpoints in WebBrain's frozen planner benchmark LFM 2.5 230M and 350M both completed the frozen WebBrain planner run without transport errors. The raw tiny checkpoints are not good browser planners yet, but their failure shape is useful fine-tuning data. We ran two tiny raw Liquid LFM 2.5 checkpoints, 230M and 350M, through WebBrain's frozen 100-case browser-agent first-tool benchmark on a local OpenAI-compatible endpoint. June 26, 2026 2026-06-26 WebBrain LFM 2.5 Liquid AI tiny language model browser agent tool calling planner benchmark Qwen 3.5">
<div class="post-meta">June 26, 2026 · 5 min read</div>
<div class="post-title">Tiny raw LFM 2.5 checkpoints in WebBrain's frozen planner benchmark</div>
<div class="post-excerpt">LFM 2.5 230M and 350M both completed the frozen WebBrain planner run without transport errors. The raw tiny checkpoints are not good browser planners yet, but their failure shape is useful fine-tuning data.</div>
</a>
<a href="/blog/minimax-m3-webbrain-cloud-tool-calling" class="post-card" data-search="MiniMax M3 and WebBrain Compass 1.0 enter the frozen WebBrain planner benchmark MiniMax M3 and WebBrain Compass 1.0 enter the frozen WebBrain planner benchmark MiniMax M3 landed at 75% Sonnet first-tool alignment, below the older MiniMax M2.7 run. WebBrain Compass 1.0 reached 73% alignment with a much lower run cost, but higher latency in this test path. We ran MiniMax M3 through OpenRouter and WebBrain Compass 1.0 through api.webbrain.one against WebBrain's frozen 100-case browser-agent first-tool benchmark, then updated the Sonnet 4.6 comparison table. June 21, 2026 2026-06-21 WebBrain MiniMax M3 WebBrain Compass OpenRouter browser agent tool calling Claude Sonnet LLM benchmark">
<div class="post-meta">June 21, 2026 · 6 min read</div>
<div class="post-title">MiniMax M3 and WebBrain Compass 1.0 enter the frozen WebBrain planner benchmark</div>
<div class="post-excerpt">MiniMax M3 landed at 75% Sonnet first-tool alignment, below the older MiniMax M2.7 run. WebBrain Compass 1.0 reached 73% alignment with a much lower run cost, but higher latency in this test path.</div>
</a>
<a href="/blog/gemma-4-12b-qat-planner-benchmark" class="post-card" data-search="Gemma 4 12B QAT is fast enough for WebBrain, but Qwen 3.5 9B still routes a little better Gemma 4 12B QAT is fast enough for WebBrain, but Qwen 3.5 9B still routes a little better Gemma 4 12B QAT lands as a very fast local WebBrain planner: 92/100 parsed calls, 14% exact, 33% tool-name match, 0.43s median per request, and a working but imperfect vision probe. We tested google/gemma-4-12B-it-qat-w4a16-ct through vLLM against WebBrain's frozen first-tool-call browser-agent harness at concurrency 8, then ran the same vision probe used in earlier local-model posts. June 20, 2026 2026-06-20 WebBrain Gemma 4 Gemma 4 12B QAT vLLM local LLM browser agent tool calling Qwen 3.5 9B">
<div class="post-meta">June 20, 2026 · 6 min read</div>
<div class="post-title">Gemma 4 12B QAT is fast enough for WebBrain, but Qwen 3.5 9B still routes a little better</div>
<div class="post-excerpt">Gemma 4 12B QAT lands as a very fast local WebBrain planner: 92/100 parsed calls, 14% exact, 33% tool-name match, 0.43s median per request, and a working but imperfect vision probe.</div>
</a>
<a href="/blog/gemma-4-31b-qat-planner-benchmark" class="post-card" data-search="Gemma 4 31B QAT quietly becomes the best local Gemma planner we have tested Gemma 4 31B QAT quietly becomes the best local Gemma planner we have tested Gemma 4 31B QAT is not branded like a new generation, but in WebBrain's local planner bench it behaves like a meaningful upgrade: 95/100 parsed calls, 19% exact, 37% tool-name match, and 0.55s median latency. We tested google/gemma-4-31B-it-qat-w4a16-ct through vLLM against WebBrain's frozen first-tool-call browser-agent harness. It improves over the older Gemma 4 31B int4 run and narrowly edges Qwen 3.6 27B on strict first-action quality while running much faster. June 20, 2026 2026-06-20 WebBrain Gemma 4 Gemma 4 31B QAT vLLM local LLM browser agent tool calling Qwen 3.6 27B">
<div class="post-meta">June 20, 2026 · 7 min read</div>
<div class="post-title">Gemma 4 31B QAT quietly becomes the best local Gemma planner we have tested</div>
<div class="post-excerpt">Gemma 4 31B QAT is not branded like a new generation, but in WebBrain's local planner bench it behaves like a meaningful upgrade: 95/100 parsed calls, 19% exact, 37% tool-name match, and 0.55s median latency.</div>
</a>
<a href="/blog/local-planner-q4-june-2026" class="post-card" data-search="DiffusionGemma hits 0.35s median in the WebBrain local planner bench DiffusionGemma hits 0.35s median in the WebBrain local planner bench Gemma 4 12B Coder, North Mini Code, and DiffusionGemma all completed the frozen legacy tool-call bench through different serving paths. DiffusionGemma was the speed surprise under vLLM and also handled the vision probe, but its diffusion-style generation still needs more WebBrain-specific reliability work. VibeThinker confirmed its own model-card warning: it is not a browser-agent tool-calling model. DiffusionGemma, Gemma 4 12B Coder, Cohere North-Mini-Code, and VibeThinker-3B in WebBrain's frozen first-tool-call browser-agent harness. June 19, 2026 2026-06-19 WebBrain local LLM browser agent tool calling Gemma 4 12B Coder North-Mini-Code DiffusionGemma VibeThinker-3B llama.cpp vLLM">
<div class="post-meta">June 19, 2026 · 9 min read</div>
<div class="post-title">DiffusionGemma hits 0.35s median in the WebBrain local planner bench</div>
<div class="post-excerpt">Gemma 4 12B Coder, North Mini Code, and DiffusionGemma all completed the frozen legacy tool-call bench through different serving paths. DiffusionGemma was the speed surprise under vLLM and also handled the vision probe, but its diffusion-style generation still needs more WebBrain-specific reliability work. VibeThinker confirmed its own model-card warning: it is not a browser-agent tool-calling model.</div>
</a>
<a href="/blog/webbrain-cloud-local-model-benchmarks" class="post-card" data-search="WebBrain Compass is live, and these are the local models we are benchmarking next WebBrain Compass is live, and these are the local models we are benchmarking next WebBrain Compass is live in the latest main branch, which means you can try WebBrain without a local LLM or API access. It is request-limited, but useful for a first look. We are also lining up the next local-model benchmark set by hardware band: 4-12GB, 12-24GB, and 24-64GB VRAM. WebBrain Compass is now live in the latest main branch, so you can try WebBrain without running a local LLM or bringing your own API key. The hosted route is request-limited, but it is enough to see how the browser agent feels. Next up: benchmarks for VibeThinker-3B, Gemma 4 12B Coder, DiffusionGemma 26B, and Cohere North-Mini-Code 30B across practical local-VRAM bands. June 19, 2026 2026-06-19 WebBrain Compass WebBrain local LLM browser agent VibeThinker-3B Gemma 4 12B Coder DiffusionGemma Cohere North-Mini-Code local model benchmark VRAM">
<div class="post-meta">June 19, 2026 · 5 min read</div>
<div class="post-title">WebBrain Compass is live, and these are the local models we are benchmarking next</div>
<div class="post-excerpt">WebBrain Compass is live in the latest main branch, which means you can try WebBrain without a local LLM or API access. It is request-limited, but useful for a first look. We are also lining up the next local-model benchmark set by hardware band: 4-12GB, 12-24GB, and 24-64GB VRAM.</div>
</a>
<a href="/blog/molmo2-8b-tool-calling" class="post-card" data-search="Molmo2-8B is truly open. Our current serving path could not give it a fair browser-tool test. Molmo2-8B is truly open. Our current serving path could not give it a fair browser-tool test. Molmo2-8B deserves praise for being open source in the meaningful sense: weights, data, recipe, and no closed-VLM distillation dependency. We could not run a fair native OpenAI-tools comparison through the current LM Studio Molmo path: structured tools failed at prompt rendering, and the fallback text-call run produced only 2 parsed tool calls. We tried to run Molmo2-8B through WebBrain's frozen 100-case browser-agent tool-calling benchmark. The open-source story is excellent: open weights, data, and recipe. The caveat: LM Studio's Molmo path rejected native structured tools, so the only successful run used a text-call compatibility shim. June 12, 2026 2026-06-12 Molmo2 Molmo2-8B Ai2 Allen Institute for AI open source AI open weights open data browser agent tool calling WebBrain LM Studio local LLM">
<div class="post-meta">June 12, 2026 · 8 min read</div>
<div class="post-title">Molmo2-8B is truly open. Our current serving path could not give it a fair browser-tool test.</div>
<div class="post-excerpt">Molmo2-8B deserves praise for being open source in the meaningful sense: weights, data, recipe, and no closed-VLM distillation dependency. We could not run a fair native OpenAI-tools comparison through the current LM Studio Molmo path: structured tools failed at prompt rendering, and the fallback text-call run produced only 2 parsed tool calls.</div>
</a>
<a href="/blog/agent-security-model" class="post-card" data-search="How WebBrain keeps a local AI agent from getting hijacked by the page it's reading How WebBrain keeps a local AI agent from getting hijacked by the page it's reading When an AI agent can act on a page as the logged-in user, the page itself becomes an attacker. We walk through WebBrain's layered defense — untrusted-content quarantine, a language-agnostic permission gate, UI-first actions — and share what our adversarial tests revealed: big models resist injection on their own, small local models need the guardrails, and the guardrails are what flip a confused model from relaying an attacker's instruction to flagging it. An act-mode browser agent can click, type, and submit as you — which means a malicious page that confuses the model is a real attack. Here's WebBrain's security model, the adversarial tests behind it, and what we found about model size and prompt injection. June 2, 2026 2026-06-02 prompt injection browser agent security AI agent safety local LLM untrusted content WebBrain accessibility tree Manifest V3">
<div class="post-meta">June 2, 2026 · 9 min read</div>
<div class="post-title">How WebBrain keeps a local AI agent from getting hijacked by the page it's reading</div>
<div class="post-excerpt">When an AI agent can act on a page as the logged-in user, the page itself becomes an attacker. We walk through WebBrain's layered defense — untrusted-content quarantine, a language-agnostic permission gate, UI-first actions — and share what our adversarial tests revealed: big models resist injection on their own, small local models need the guardrails, and the guardrails are what flip a confused model from relaying an attacker's instruction to flagging it.</div>
</a>
<a href="/blog/liquid-lfm25-tool-calling" class="post-card" data-search="Liquid LFM 2.5-8B-A1B on browser tool calling: where the new on-device MoE fits Liquid LFM 2.5-8B-A1B on browser tool calling: where the new on-device MoE fits Liquid AI's just-shipped LFM 2.5-8B-A1B (8.3B total / 1.5B active sparse MoE) joins our browser-agent benchmark. It beats both small Qwens and runs within reach of Gemma 4-E2B against Claude Sonnet 4.6 — but for the small-model distillation slot, plain Apache 2.0 still wins over LFM's $10M-revenue-cliff license. Also: a tool-schema drift bug we caught and the freeze fix we shipped. We added Liquid AI's LFM 2.5-8B-A1B (8.3B total / 1.5B active MoE) to our browser-agent tool-calling benchmark. Against Claude Sonnet 4.6 it beats both small Qwens but loses to Gemma 4-E2B. Plus: a tool-schema drift bug that changes how we score everything, and why the license asymmetry (LFM's $10M revenue cliff vs Gemma 4's plain Apache 2.0) tilts the distillation call to Gemma. May 31, 2026 2026-05-31 LFM 2.5 LFM2.5-8B-A1B Liquid AI on-device LLM MoE tool calling browser agent Claude Sonnet 4.6 WebBrain Liquid Foundation Model sparse mixture of experts">
<div class="post-meta">May 31, 2026 · 7 min read</div>
<div class="post-title">Liquid LFM 2.5-8B-A1B on browser tool calling: where the new on-device MoE fits</div>
<div class="post-excerpt">Liquid AI's just-shipped LFM 2.5-8B-A1B (8.3B total / 1.5B active sparse MoE) joins our browser-agent benchmark. It beats both small Qwens and runs within reach of Gemma 4-E2B against Claude Sonnet 4.6 — but for the small-model distillation slot, plain Apache 2.0 still wins over LFM's $10M-revenue-cliff license. Also: a tool-schema drift bug we caught and the freeze fix we shipped.</div>
</a>
<a href="/blog/llm-tool-calling-benchmark" class="post-card" data-search="13 LLMs, 100 browser tasks, two baselines: which model actually picks the right tool? 13 LLMs, 100 browser tasks, two baselines: which model actually picks the right tool? We benchmarked 13 local and API models on 100 real browser-agent tool-calling tasks against both consensus voting and Claude Sonnet 4.6. The consensus winner (Qwen 3.6-35B-A3B at 94%) isn't the Sonnet-match winner (Qwen 3.6-27B at 77%) — and that gap tells you something useful about what "correct" means for tool calling. We benchmarked 13 local and API LLMs on 100 real browser-agent tool-calling tasks against both consensus voting and Claude Sonnet 4.6. The winner depends on which question you ask. May 23, 2026 2026-05-23 LLM benchmark tool calling browser agent Qwen 3.6 Claude Sonnet consensus voting WebBrain self-hosted LLM MoE Gemma 4">
<div class="post-meta">May 23, 2026 · 8 min read</div>
<div class="post-title">13 LLMs, 100 browser tasks, two baselines: which model actually picks the right tool?</div>
<div class="post-excerpt">We benchmarked 13 local and API models on 100 real browser-agent tool-calling tasks against both consensus voting and Claude Sonnet 4.6. The consensus winner (Qwen 3.6-35B-A3B at 94%) isn't the Sonnet-match winner (Qwen 3.6-27B at 77%) — and that gap tells you something useful about what "correct" means for tool calling.</div>
</a>
<a href="/blog/gemma4-turkish-pruning" class="post-card" data-search="Pruning Gemma 4 26B-A4B to run on small GPUs Pruning Gemma 4 26B-A4B for small GPUs: Turkish-first, language-agnostic MoE surgery Router hooks + expert activation telemetry + surgical long-tail removal + brief LoRA heal. Early run: 128→101 experts/layer, 26B→21B params, ~11 GB at 4-bit GGUF, with solid Turkish fluency and code performance. A practical Gemma 4 26B-A4B expert-pruning run: remove low-activation experts for Turkish+English workloads, then recover quality with a short LoRA heal. May 19, 2026 2026-05-19 Gemma 4 MoE pruning expert pruning Turkish LLM GGUF IQ4_XS LoRA small GPU WebBrain">
<div class="post-meta">May 19, 2026 · 4 min read</div>
<div class="post-title">Pruning Gemma 4 26B-A4B for small GPUs: Turkish-first, language-agnostic MoE surgery</div>
<div class="post-excerpt">Router hooks + expert activation telemetry + surgical long-tail removal + brief LoRA heal. Early run: 128→101 experts/layer, 26B→21B params, ~11 GB at 4-bit GGUF, with solid Turkish fluency and code performance.</div>
</a>
<a href="/blog/vision-shootout-round-4" class="post-card" data-search="Round 4: Qwen 3.5-9B-int4 punches above its weight — when 9B int4 beats 308B IQ3_S on affordance Round 4: Qwen 3.5-9B-int4 punches above its weight — when 9B int4 beats 308B IQ3_S on affordance A 9B int4 Qwen 3.5 on vLLM classifies the email-chip dropdown explicitly — something the 308B MiMo V2.5 at IQ3_S and the larger Qwen 3.6-35B-A3B both missed. Cheapest image tokens after Gemma. The catch: it loses the red-border visual cue and joins the "Unknowns: None" club. Suddenly the most interesting option for ≤8 GB VRAM. Plus an updated routing-policy table by VRAM bracket. Round 4 of the vision-model shootout for browser agents. A 9B int4 Qwen 3.5 model classifies the email-chip dropdown explicitly — something the 308B MiMo IQ3_S and the larger Qwen 3.6-A3B both missed. Cheapest image tokens after Gemma. The catch: it loses the red-border visual cue and joins the 'Unknowns: None' club. May 7, 2026 2026-05-07 Qwen 3.5 9B AutoRound int4 vLLM vision vision model benchmark browser agent affordance classification MiMo V2.5 vs Qwen low VRAM vision model dropdown classification calibrated uncertainty Intel AutoRound llama.cpp vs vLLM">
<div class="post-meta">May 7, 2026 · 6 min read</div>
<div class="post-title">Round 4: Qwen 3.5-9B-int4 punches above its weight — when 9B int4 beats 308B IQ3_S on affordance</div>
<div class="post-excerpt">A 9B int4 Qwen 3.5 on vLLM classifies the email-chip dropdown explicitly — something the 308B MiMo V2.5 at IQ3_S and the larger Qwen 3.6-35B-A3B both missed. Cheapest image tokens after Gemma. The catch: it loses the red-border visual cue and joins the "Unknowns: None" club. Suddenly the most interesting option for ≤8 GB VRAM. Plus an updated routing-policy table by VRAM bracket.</div>
</a>
<a href="/blog/vision-shootout-round-3" class="post-card" data-search="Round 3: Xiaomi MiMo V2.5 enters the vision shootout — and joins the "Unknowns: None" club Round 3: Xiaomi MiMo V2.5 enters the vision shootout — and joins the "Unknowns: None" club The empirical follow-up to last week's MiMo speculation post. Same probe, same Google sign-in screen, same prompt. MiMo at IQ3_S nails OCR and state extraction in the Qwen 3.6 tier — but joins Nemotron and Gemma in writing "Unknowns: None" instead of flagging the red-border ambiguity. Token cost ties the most expensive bucket; latency is in a different class. Plus a probe upgrade so big reasoning models stop tripping the default fetch headers timeout. A third round of vision-model benchmarking for browser agents. Xiaomi's MiMo-V2.5 — the "looks promising" model from our research notes — sits down at the same probe with the same Google sign-in screen. Strong OCR, expensive image tokens, no calibrated uncertainty. May 7, 2026 2026-05-07 MiMo V2.5 Xiaomi MiMo vision model benchmark browser agent llama.cpp vision IQ3_S quantization calibrated uncertainty Qwen 3.6 vs MiMo multimodal model omni-modal reasoning model undici headers timeout vision probe">
<div class="post-meta">May 7, 2026 · 7 min read</div>
<div class="post-title">Round 3: Xiaomi MiMo V2.5 enters the vision shootout — and joins the "Unknowns: None" club</div>
<div class="post-excerpt">The empirical follow-up to last week's MiMo speculation post. Same probe, same Google sign-in screen, same prompt. MiMo at IQ3_S nails OCR and state extraction in the Qwen 3.6 tier — but joins Nemotron and Gemma in writing "Unknowns: None" instead of flagging the red-border ambiguity. Token cost ties the most expensive bucket; latency is in a different class. Plus a probe upgrade so big reasoning models stop tripping the default fetch headers timeout.</div>
</a>
<a href="/blog/mimo-v25-pro-vs-flash" class="post-card" data-search="Xiaomi MiMo V2.5 Pro vs "V2.5 Flash": should WebBrain add both? Xiaomi MiMo V2.5 Pro vs "V2.5 Flash": should WebBrain add both? Research notes on Xiaomi's newly released MiMo V2.5 series and why multimodal Pro+Flash-style routing may outperform text-only stacks for browser-agent workloads, while Qwen 3.6 still leads value on many pure-text tasks. Research notes on Xiaomi's MiMo V2.5 Pro and the V2.5/Flash-style tier: why these multimodal models look promising for WebBrain compared with text-only alternatives, and how they slot into our routing policy alongside Qwen 3.6. April 30, 2026 2026-04-30 Xiaomi MiMo MiMo V2.5 Pro MiMo V2.5 Flash multimodal model routing browser agent WebBrain Qwen 3.6 DeepSeek-V4 MiniMax M2.7 Nemotron-3-Nano-Omni OpenRouter model routing">
<div class="post-meta">April 30, 2026 · 6 min read</div>
<div class="post-title">Xiaomi MiMo V2.5 Pro vs "V2.5 Flash": should WebBrain add both?</div>
<div class="post-excerpt">Research notes on Xiaomi's newly released MiMo V2.5 series and why multimodal Pro+Flash-style routing may outperform text-only stacks for browser-agent workloads, while Qwen 3.6 still leads value on many pure-text tasks.</div>
</a>
<a href="/blog/vision-shootout-round-2" class="post-card" data-search="Round 2: Nemotron Omni 30B vs Qwen 3.6 — does cheaper image tokens beat calibrated uncertainty? Round 2: Nemotron Omni 30B vs Qwen 3.6 — does cheaper image tokens beat calibrated uncertainty? A second round of vision-model benchmarking for browser agents. NVIDIA's Nemotron Omni 30B-A3B-Reasoning is 17% cheaper per image and classifies inputs better than Qwen 3.6-35B-A3B — but loses on calibrated uncertainty, and is English-only. Plus a head-to-head with the dense Qwen 3.6-27B that explains why MoE is the right architecture for self-hosted vision. A second round of vision-model benchmarking for browser agents. NVIDIA's Nemotron Omni 30B-A3B-Reasoning and the dense Qwen 3.6-27B both enter — and one of them is 17% cheaper per image. We measure what changes for a real browser-agent workload, and what doesn't. April 29, 2026 2026-04-29 Nemotron Omni 30B Qwen 3.6 35B Qwen 3.6 27B vision model benchmark browser agent MoE vision model calibrated uncertainty vision encoder cost NVFP4 llama.cpp vLLM A3B reasoning model">
<div class="post-meta">April 29, 2026 · 8 min read</div>
<div class="post-title">Round 2: Nemotron Omni 30B vs Qwen 3.6 — does cheaper image tokens beat calibrated uncertainty?</div>
<div class="post-excerpt">A second round of vision-model benchmarking for browser agents. NVIDIA's Nemotron Omni 30B-A3B-Reasoning is 17% cheaper per image and classifies inputs better than Qwen 3.6-35B-A3B — but loses on calibrated uncertainty, and is English-only. Plus a head-to-head with the dense Qwen 3.6-27B that explains why MoE is the right architecture for self-hosted vision.</div>
</a>
<a href="/blog/vision-model-shootout" class="post-card" data-search="Four vision models, one screenshot: which one is actually worth running locally for a browser agent? Four vision models, one screenshot: which one is actually worth running locally for a browser agent? We fed the same Google sign-in page through Gemma 4-E2B, Gemma 4-31B, Qwen3.5-27B, and Qwen3.6-35B-A3B using the exact system prompt WebBrain's vision sub-call ships with. The spread on OCR accuracy, latency, and token cost is wider than you'd expect — and one model quietly changed our mind about which architecture to reach for. We benchmarked Gemma 4-E2B, Gemma 4-31B, Qwen3.5-27B, and Qwen3.6-35B-A3B on the same browser screenshot using WebBrain's exact vision-sub-call prompt. Here's what the numbers look like and why one model changed our default recommendation. April 21, 2026 2026-04-21 vision model benchmark open source browser agent Qwen3.6 Gemma 4 llama.cpp vision local vision model Qwen3.5 27B browser automation OCR">
<div class="post-meta">April 21, 2026 · 7 min read</div>
<div class="post-title">Four vision models, one screenshot: which one is actually worth running locally for a browser agent?</div>
<div class="post-excerpt">We fed the same Google sign-in page through Gemma 4-E2B, Gemma 4-31B, Qwen3.5-27B, and Qwen3.6-35B-A3B using the exact system prompt WebBrain's vision sub-call ships with. The spread on OCR accuracy, latency, and token cost is wider than you'd expect — and one model quietly changed our mind about which architecture to reach for.</div>
</a>
</div>
<div class="empty-state" id="empty-state" hidden>
<strong>No notes match “<span id="empty-query"></span>”</strong>
Try a model name, benchmark, or tool instead.
<div><button class="empty-clear" id="empty-clear" type="button">Clear search</button></div>
</div>
<div class="pagination" id="pagination" role="navigation" aria-label="Blog pages" hidden>
<button class="page-button page-direction" id="page-previous" type="button">Previous</button>
<div class="page-numbers" id="page-numbers"></div>
<button class="page-button page-direction" id="page-next" type="button">Next</button>
</div>
</main>
<!-- Site footer: start -->
<footer class="wb-footer">
<div class="wb-footer-inner">
<div class="wb-footer-brand">
<p class="wb-footer-tagline">On a mission to make the web smarter.</p>
<p class="wb-footer-copy">© 2026 WebBrain. Built by <a href="https://emresokullu.com" target="_blank">Emre Sokullu</a> & <a href="https://github.com/webbrain-one/webbrain/graphs/contributors" target="_blank">contributors</a>.</p>
</div>
<div class="wb-footer-row">
<div class="wb-footer-links">
<a href="/">Browser extension</a>
<a href="/companion/">Desktop Companion</a>
<a href="/research/">Research</a>
<a href="/docs/">Docs</a>
<a href="/docs/faq/">FAQ</a>
<a href="/blog/">Blog</a>
<a href="/privacy">Privacy Policy</a>
<a href="https://github.com/webbrain-one/webbrain/issues" target="_blank" rel="noopener">Report Bug</a>
<a href="https://github.com/webbrain-one/webbrain/blob/main/LICENSE" target="_blank" rel="noopener">GPL-3.0 License</a>
<a href="https://pressreleases.online/org/webbrain.one" target="_blank" rel="noopener">Press</a>
<a href="https://api.webbrain.one/account" target="_blank" rel="noopener">Compass Subscription</a>
</div>
<div class="wb-footer-social">
<a href="https://mastoturk.org/@webbrain" target="_blank" rel="me noopener" aria-label="WebBrain on Mastodon" title="WebBrain on Mastodon" class="wb-footer-icon"><svg viewBox="0 0 24 24" fill="currentColor" width="15" height="15" aria-hidden="true"><path d="M21.327 8.566c0-5.035-3.298-6.511-3.298-6.511C16.366 1.292 13.508.971 12.04.947h-.036c-1.468.024-4.326.345-5.989 1.108 0 0-3.298 1.476-3.298 6.511 0 1.153-.022 2.535.014 4.001.12 4.935.907 9.799 5.478 11.007 2.106.557 3.916.674 5.374.594 2.645-.147 4.128-.943 4.128-.943l-.087-1.918s-1.888.595-4.012.523c-2.104-.072-4.326-.227-4.668-2.809a5.289 5.289 0 0 1-.047-.724s2.066.505 4.686.625c1.603.073 3.107-.094 4.632-.276 2.925-.349 5.474-2.149 5.794-3.793.503-2.591.462-6.287.462-6.287ZM8.545 15.023H6.09V8.996c0-2.534 1.627-3.829 3.424-3.829 1.963 0 2.909 1.268 3.35 2.257.44-.989 1.386-2.257 3.349-2.257 1.797 0 3.424 1.295 3.424 3.829v6.027h-2.454V9.174c0-1.233-.519-1.859-1.557-1.859-1.148 0-1.724.742-1.724 2.207v3.203h-2.454V9.522c0-1.465-.576-2.207-1.724-2.207-1.038 0-1.557.626-1.557 1.859v5.849Z"/></svg></a>
<a href="https://bsky.app/profile/webbrain-one.bsky.social" target="_blank" rel="me noopener" aria-label="WebBrain on Bluesky" title="WebBrain on Bluesky" class="wb-footer-icon"><svg viewBox="0 0 24 24" fill="currentColor" width="15" height="15" aria-hidden="true"><path d="M12 10.8c-1.1-2.1-4-6-6.8-8C2.6.9 1.6 1.3.9 1.6.1 1.9 0 3.1 0 3.8c0 .7.4 5.7.6 6.5.8 2.7 3.7 3.7 5.6 3.4-3.5.6-6.5 2-2.5 6.5 4.3 4.5 5.9-1 6.3-2.8.1-.5.2-.8.3-.8.1 0 .2.3.3.8.4 1.8 2 7.3 6.3 2.8 4-4.5 1-5.9-2.5-6.5 1.9.3 4.8-.7 5.6-3.4.2-.8.6-5.8.6-6.5 0-.7-.1-1.9-.9-2.2-.7-.3-1.7-.6-4.3 1.2-2.8 2-5.7 5.9-6.8 8Z"/></svg></a>
<a href="https://discord.gg/cgC325ssfw" target="_blank" rel="noopener" aria-label="WebBrain on Discord" title="WebBrain on Discord" class="wb-footer-icon"><svg viewBox="0 0 16 16" fill="currentColor" width="15" height="15" aria-hidden="true"><path d="M13.545 2.907a13.227 13.227 0 0 0-3.257-1.011c.05.069.098.137.142.207a12.471 12.471 0 0 0-3.616 0c.044-.07.093-.138.143-.207A13.207 13.207 0 0 0 3.7 2.907 13.136 13.136 0 0 0 .931 13.362a13.667 13.667 0 0 0 4.022 2.018c.33-.442.623-.907.87-1.396a8.95 8.95 0 0 1-1.37-.652c.114-.083.225-.171.33-.269a9.6 9.6 0 0 0 8.243 0c.108.09.218.18.33.269a8.923 8.923 0 0 1-1.37.652c.248.489.54.954.87 1.396a13.667 13.667 0 0 0 4.022-2.018 13.136 13.136 0 0 0-2.769-10.455zM5.106 11.275c-.788 0-1.438-.72-1.438-1.608 0-.889.625-1.608 1.438-1.608.818 0 1.451.726 1.438 1.608 0 .888-.625 1.608-1.438 1.608zm5.788 0c-.788 0-1.438-.72-1.438-1.608 0-.889.625-1.608 1.438-1.608.818 0 1.451.726 1.438 1.608 0 .888-.625 1.608-1.438 1.608z"/></svg></a>
<a href="https://huggingface.co/webbrain-one" target="_blank" rel="me noopener" aria-label="WebBrain on Hugging Face" title="WebBrain on Hugging Face" class="wb-footer-icon"><svg viewBox="0 0 24 24" fill="currentColor" width="15" height="15" aria-hidden="true"><path d="M12.025 1.13c-5.77 0-10.449 4.647-10.449 10.378 0 1.112.178 2.181.503 3.185.064-.222.203-.444.416-.577a.96.96 0 0 1 .524-.15c.293 0 .584.124.84.284.278.173.48.408.71.694.226.282.458.611.684.951v-.014c.017-.324.106-.622.264-.874s.403-.487.762-.543c.3-.047.596.06.787.203s.31.313.4.467c.15.257.212.468.233.542.01.026.653 1.552 1.657 2.54.616.605 1.01 1.223 1.082 1.912.055.537-.096 1.059-.38 1.572.637.121 1.294.187 1.967.187.657 0 1.298-.063 1.921-.178-.287-.517-.44-1.041-.384-1.581.07-.69.465-1.307 1.081-1.913 1.004-.987 1.647-2.513 1.657-2.539.021-.074.083-.285.233-.542.09-.154.208-.323.4-.467a1.08 1.08 0 0 1 .787-.203c.359.056.604.29.762.543s.247.55.265.874v.015c.225-.34.457-.67.683-.952.23-.286.432-.52.71-.694.257-.16.547-.284.84-.285a.97.97 0 0 1 .524.151c.228.143.373.388.43.625l.006.04a10.3 10.3 0 0 0 .534-3.273c0-5.731-4.678-10.378-10.449-10.378M8.327 6.583a1.5 1.5 0 0 1 .713.174 1.487 1.487 0 0 1 .617 2.013c-.183.343-.762-.214-1.102-.094-.38.134-.532.914-.917.71a1.487 1.487 0 0 1 .69-2.803m7.486 0a1.487 1.487 0 0 1 .689 2.803c-.385.204-.536-.576-.916-.71-.34-.12-.92.437-1.103.094a1.487 1.487 0 0 1 .617-2.013 1.5 1.5 0 0 1 .713-.174m-10.68 1.55a.96.96 0 1 1 0 1.921.96.96 0 0 1 0-1.92m13.838 0a.96.96 0 1 1 0 1.92.96.96 0 0 1 0-1.92M8.489 11.458c.588.01 1.965 1.157 3.572 1.164 1.607-.007 2.984-1.155 3.572-1.164.196-.003.305.12.305.454 0 .886-.424 2.328-1.563 3.202-.22-.756-1.396-1.366-1.63-1.32q-.011.001-.02.006l-.044.026-.01.008-.03.024q-.018.017-.035.036l-.032.04a1 1 0 0 0-.058.09l-.014.025q-.049.088-.11.19a1 1 0 0 1-.083.116 1.2 1.2 0 0 1-.173.18q-.035.029-.075.058a1.3 1.3 0 0 1-.251-.243 1 1 0 0 1-.076-.107c-.124-.193-.177-.363-.337-.444-.034-.016-.104-.008-.2.022q-.094.03-.216.087-.06.028-.125.063l-.13.074q-.067.04-.136.086a3 3 0 0 0-.135.096 3 3 0 0 0-.26.219 2 2 0 0 0-.12.121 2 2 0 0 0-.106.128l-.002.002a2 2 0 0 0-.09.132l-.001.001a1.2 1.2 0 0 0-.105.212q-.013.036-.024.073c-1.139-.875-1.563-2.317-1.563-3.203 0-.334.109-.457.305-.454m.836 10.354c.824-1.19.766-2.082-.365-3.194-1.13-1.112-1.789-2.738-1.789-2.738s-.246-.945-.806-.858-.97 1.499.202 2.362c1.173.864-.233 1.45-.685.64-.45-.812-1.683-2.896-2.322-3.295s-1.089-.175-.938.647 2.822 2.813 2.562 3.244-1.176-.506-1.176-.506-2.866-2.567-3.49-1.898.473 1.23 2.037 2.16c1.564.932 1.686 1.178 1.464 1.53s-3.675-2.511-4-1.297c-.323 1.214 3.524 1.567 3.287 2.405-.238.839-2.71-1.587-3.216-.642-.506.946 3.49 2.056 3.522 2.064 1.29.33 4.568 1.028 5.713-.624m5.349 0c-.824-1.19-.766-2.082.365-3.194 1.13-1.112 1.789-2.738 1.789-2.738s.246-.945.806-.858.97 1.499-.202 2.362c-1.173.864.233 1.45.685.64.451-.812 1.683-2.896 2.322-3.295s1.089-.175.938.647-2.822 2.813-2.562 3.244 1.176-.506 1.176-.506 2.866-2.567 3.49-1.898-.473 1.23-2.037 2.16c-1.564.932-1.686 1.178-1.464 1.53s3.675-2.511 4-1.297c.323 1.214-3.524 1.567-3.287 2.405.238.839 2.71-1.587 3.216-.642.506.946-3.49 2.056-3.522 2.064-1.29.33-4.568 1.028-5.713-.624"/></svg></a>
</div>
</div>
</div>
</footer>
<!-- Site footer: end -->
<script>
(function () {
const btn = document.getElementById('theme-toggle');
if (!btn) return;
const root = document.documentElement;
function setTheme(theme, persist) {
root.setAttribute('data-theme', theme);
if (persist) {
try { localStorage.setItem('webbrain-theme', theme); } catch (_) {}
}
}
btn.addEventListener('click', function () {
const next = root.getAttribute('data-theme') === 'light' ? 'dark' : 'light';
setTheme(next, true);
});
const mq = window.matchMedia('(prefers-color-scheme: light)');
function onMQ(e) {
try {
if (localStorage.getItem('webbrain-theme')) return;
} catch (_) {}
setTheme(e.matches ? 'light' : 'dark', false);
}
if (typeof mq.addEventListener === 'function') mq.addEventListener('change', onMQ);
else if (typeof mq.addListener === 'function') mq.addListener(onMQ);
})();
</script>
<script>
(function () {
const LOCALES = ["en","es","fr","tr","zh","ru","uk","ar","ja","ko","id","th","ms","tl","he"];
const sel = document.getElementById('lang-dropdown');
if (!sel) return;
sel.value = sel.dataset.current || 'en';
sel.addEventListener('change', function () {
const target = sel.value;
if (!LOCALES.includes(target)) return;
window.location.href = target === 'en' ? '/' : '/' + target + '/';
});
})();
</script>
<script>
(function () {
const PAGE_SIZE = 8;
const searchInput = document.getElementById('blog-search');
const clearButton = document.getElementById('search-clear');
const emptyClearButton = document.getElementById('empty-clear');
const resultsCount = document.getElementById('results-count');
const pageStatus = document.getElementById('page-status');
const emptyState = document.getElementById('empty-state');
const emptyQuery = document.getElementById('empty-query');
const pagination = document.getElementById('pagination');
const pageNumbers = document.getElementById('page-numbers');
const previousButton = document.getElementById('page-previous');
const nextButton = document.getElementById('page-next');
const tools = document.getElementById('blog-tools');
const cards = Array.from(document.querySelectorAll('.post-card'));
if (!searchInput || !cards.length) return;
let query = '';
let currentPage = 1;
function normalize(value) {
return String(value || '')
.normalize('NFD')
.replace(/[\u0300-\u036f]/g, '')
.toLocaleLowerCase();
}
function readUrl() {
const params = new URLSearchParams(window.location.search);
query = (params.get('q') || '').trim();
const requestedPage = Number.parseInt(params.get('page') || '1', 10);
currentPage = Number.isFinite(requestedPage) && requestedPage > 0 ? requestedPage : 1;
searchInput.value = query;
}
function writeUrl(mode) {
if (!mode) return;
const url = new URL(window.location.href);
if (query) url.searchParams.set('q', query);
else url.searchParams.delete('q');
if (currentPage > 1) url.searchParams.set('page', String(currentPage));
else url.searchParams.delete('page');
const target = url.pathname + (url.searchParams.toString() ? '?' + url.searchParams.toString() : '') + url.hash;
window.history[mode + 'State']({}, '', target);
}
function pageSequence(total, active) {
if (total <= 7) return Array.from({ length: total }, function (_, index) { return index + 1; });
const candidates = [1, total, active - 2, active - 1, active, active + 1, active + 2]
.filter(function (value) { return value >= 1 && value <= total; });
const pages = Array.from(new Set(candidates)).sort(function (a, b) { return a - b; });
const sequence = [];
pages.forEach(function (value, index) {
if (index && value - pages[index - 1] > 1) sequence.push('ellipsis-' + value);
sequence.push(value);
});
return sequence;
}
function goToPage(page) {
currentPage = page;
render('push');
tools.scrollIntoView({ block: 'start' });
window.requestAnimationFrame(function () {
const active = pageNumbers.querySelector('[aria-current="page"]');
if (active) active.focus({ preventScroll: true });
});
}
function renderPageButtons(totalPages) {
pageNumbers.replaceChildren();
pageSequence(totalPages, currentPage).forEach(function (entry) {
if (typeof entry === 'string') {
const ellipsis = document.createElement('span');
ellipsis.className = 'page-ellipsis';
ellipsis.textContent = '…';
ellipsis.setAttribute('aria-hidden', 'true');
pageNumbers.appendChild(ellipsis);
return;
}
const button = document.createElement('button');
button.type = 'button';
button.className = 'page-button';
button.textContent = String(entry);
button.setAttribute('aria-label', 'Go to page ' + entry);
if (entry === currentPage) button.setAttribute('aria-current', 'page');
button.addEventListener('click', function () { goToPage(entry); });
pageNumbers.appendChild(button);
});
}
function render(historyMode) {
const terms = normalize(query).split(/\s+/).filter(Boolean);
const matches = cards.filter(function (card) {
const haystack = normalize(card.dataset.search);
return terms.every(function (term) { return haystack.includes(term); });
});
const totalPages = Math.max(1, Math.ceil(matches.length / PAGE_SIZE));
currentPage = Math.min(currentPage, totalPages);
const visible = new Set(matches.slice((currentPage - 1) * PAGE_SIZE, currentPage * PAGE_SIZE));
cards.forEach(function (card) { card.hidden = !visible.has(card); });
clearButton.hidden = !query;
emptyState.hidden = matches.length !== 0;
emptyQuery.textContent = query;
pagination.hidden = matches.length === 0 || totalPages <= 1;
previousButton.disabled = currentPage === 1;
nextButton.disabled = currentPage === totalPages;
resultsCount.textContent = query
? matches.length + (matches.length === 1 ? ' note found' : ' notes found')
: matches.length + (matches.length === 1 ? ' note' : ' notes');
pageStatus.textContent = matches.length ? 'Page ' + currentPage + ' of ' + totalPages : '';
renderPageButtons(totalPages);
writeUrl(historyMode);
}
function clearSearch() {
query = '';
currentPage = 1;
searchInput.value = '';
render('push');
searchInput.focus();
}
searchInput.addEventListener('input', function () {
query = searchInput.value.trim();
currentPage = 1;
render('replace');
});
clearButton.addEventListener('click', clearSearch);
emptyClearButton.addEventListener('click', clearSearch);
previousButton.addEventListener('click', function () { if (currentPage > 1) goToPage(currentPage - 1); });
nextButton.addEventListener('click', function () { goToPage(currentPage + 1); });
window.addEventListener('popstate', function () { readUrl(); render(); });
document.addEventListener('keydown', function (event) {
const target = event.target;
const isEditable = target instanceof HTMLElement && (
target.isContentEditable || /^(INPUT|TEXTAREA|SELECT)$/.test(target.tagName)
);
if (event.key === '/' && !isEditable && !event.metaKey && !event.ctrlKey && !event.altKey) {
event.preventDefault();
searchInput.focus();
searchInput.select();
}
if (event.key === 'Escape' && document.activeElement === searchInput) {
if (searchInput.value) clearSearch();
else searchInput.blur();
}
});
readUrl();
render('replace');
})();
</script>
</body>
</html>