Open-source AI browser agent for Chrome and Firefox (monorepo) đź§
JavaScript ai-agentai-agentsai-toolsbrowser-aibrowser-automationbrowser-extensionchrome-aichrome-devtoolschrome-extensionchrome-extensionsfirefox-addonfirefox-addonsfirefox-extensionfirefox-extensionsgemmagemma4minimaxoffline-aioffline-llmqwen
web: correct Compass Tiny v2 release naming
Emre Sokullue89e0c17caa6
diff --git a/test/llm/analysis/webbrain-compass-v2-bf16-20260910.md b/test/llm/analysis/webbrain-compass-v2-bf16-20260910.mdindex c3c347a5d..4a0f3a351 100644--- a/test/llm/analysis/webbrain-compass-v2-bf16-20260910.md+++ b/test/llm/analysis/webbrain-compass-v2-bf16-20260910.md@@ -1,4 +1,7 @@-# WebBrain Compass v2: merged BF16 routing results+# WebBrain Compass Tiny v2: merged BF16 routing results++The model's corrected release name is WebBrain Compass Tiny v2. Historical+run IDs and artifact paths below are retained unchanged for reproducibility.Completed 2026-09-10. This is the full merged BF16 candidate, not the earlierunmerged adapter and not ONNX. Base: OpenBMB MiniCPM5-2B (~2.6B), revisiondiff --git a/web/blog/index.html b/web/blog/index.htmlindex e0b18e7fc..2744c3417 100644--- a/web/blog/index.html+++ b/web/blog/index.html@@ -671,10 +671,10 @@"blogPost": [{"@type": "BlogPosting",- "headline": "WebBrain Compass v2: small model, sharper tool routing",+ "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 v2, our MiniCPM5-2B-based tool-routing model. Compare its merged BF16 results with 11 other models, including larger reference models.",+ "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",@@ -1343,10 +1343,10 @@</section><div class="post-list" id="post-list">- <a href="/blog/webbrain-compass-v2" class="post-card" data-search="WebBrain Compass v2: small model, sharper tool routing WebBrain Compass v2: small model, sharper tool routing Built on MiniCPM5-2B, Compass v2 reaches 16 exact actions and 41 tool-family matches out of 89 scenarios. A promising step toward WebBrain integration. Meet WebBrain Compass 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 v2 MiniCPM5-2B tool calling browser agent benchmark local AI">+ <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 v2: small model, sharper tool routing</div>- <div class="post-excerpt">Built on MiniCPM5-2B, Compass v2 reaches 16 exact actions and 41 tool-family matches out of 89 scenarios. A promising step toward WebBrain integration.</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>diff --git a/web/blog/posts/webbrain-compass-v2.md b/web/blog/posts/webbrain-compass-v2.mdindex 86953178c..f486a21da 100644--- a/web/blog/posts/webbrain-compass-v2.md+++ b/web/blog/posts/webbrain-compass-v2.md@@ -1,36 +1,36 @@---title: >- WebBrain Compass v2: small model, sharper tool routing+ WebBrain Compass Tiny v2: small model, sharper tool routingslug: webbrain-compass-v2sortOrder: -283date: 2026-09-10readTime: 2 min readdescription: >- Meet WebBrain Compass v2, our MiniCPM5-2B-based tool-routing model. Compare its merged BF16 results with 11 other models, including larger reference models.+ 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.excerpt: >- Built on MiniCPM5-2B, Compass v2 reaches 16 exact actions and 41 tool-family matches out of 89 scenarios. A promising step toward WebBrain integration.+ 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.titleTag: >- WebBrain Compass v2: MiniCPM5-Based Tool Routing - WebBrain Blog+ WebBrain Compass Tiny v2: MiniCPM5-Based Tool Routing - WebBrain BlogogTitle: >- Meet WebBrain Compass v2+ Meet WebBrain Compass Tiny v2ogDescription: >A small, specialized tool-routing model: 16/89 exact actions and 41/89 tool-family matches in our Compact routing benchmark.author: Emre SokulluauthorUrl: https://emresokullu.comkeywords:- - WebBrain Compass v2+ - WebBrain Compass Tiny v2- MiniCPM5-2B- tool calling- browser agent benchmark- local AIhtml: truelede: >- **Meet WebBrain Compass v2**, our new small model for choosing WebBrain browser tools and their arguments. Built on MiniCPM5-2B, it shows a substantial routing improvement over its base. We expect to incorporate it into WebBrain soon, subject to runtime validation and release checks.+ **Meet WebBrain Compass Tiny v2**, our new small model for choosing WebBrain browser tools and their arguments. Built on MiniCPM5-2B, it shows a substantial routing improvement over its base. We expect to incorporate it into WebBrain soon, subject to runtime validation and release checks.---## A focused upgrade-Compass v2 is a LoRA fine-tune of [OpenBMB's MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B), approximately 2.6B parameters. We trained it for one epoch on 7,311 reviewed examples, with 213 held out for validation. Its job is specialized tool routing—not replacing a general-purpose planner.+Compass Tiny v2 is a LoRA fine-tune of [OpenBMB's MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B), approximately 2.6B parameters. We trained it for one epoch on 7,311 reviewed examples, with 213 held out for validation. Its job is specialized tool routing—not replacing a general-purpose planner.Against the separately tested base, the merged BF16 model moves from **9 to 16 exact actions** and **21 to 41 tool-family matches**. Structured-call coverage slips from 96 to 94 out of 100.@@ -66,14 +66,14 @@ Sorted by strict exact action, then loose match. Larger Qwen models are included<thead><tr><th scope="col">Model</th><th scope="col">Parameters</th><th scope="col">Structured calls</th><th scope="col">Strict exact action</th><th scope="col">Loose tool-family match</th><th scope="col">Loose rate</th></tr></thead><tbody><tr><th scope="row">Qwen3.8-27B</th><td>27B</td><td>99<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:19.1011%"></i><b>17<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:43.8202%"></i><b>39<span class="denom"> / 89</span></b></span></td><td>43.8%</td></tr>-<tr class="compass"><th scope="row">WebBrain Compass v2<small>Merged BF16 · new</small></th><td>~2.6B</td><td>94<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:17.9775%"></i><b>16<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:46.0674%"></i><b>41<span class="denom"> / 89</span></b></span></td><td>46.1%</td></tr>+<tr class="compass"><th scope="row">WebBrain Compass Tiny v2<small>Merged BF16 · new</small></th><td>~2.6B</td><td>94<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:17.9775%"></i><b>16<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:46.0674%"></i><b>41<span class="denom"> / 89</span></b></span></td><td>46.1%</td></tr><tr><th scope="row">Qwen3.5-4B</th><td>4B</td><td>89<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:17.9775%"></i><b>16<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:40.4494%"></i><b>36<span class="denom"> / 89</span></b></span></td><td>40.4%</td></tr><tr><th scope="row">Gemma 4 E4B</th><td>7.5B</td><td>90<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:16.8539%"></i><b>15<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:48.3146%"></i><b>43<span class="denom"> / 89</span></b></span></td><td>48.3%</td></tr><tr><th scope="row">Qwen3.6-35B-A3B</th><td>35B A3B</td><td>89<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:16.8539%"></i><b>15<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:47.1910%"></i><b>42<span class="denom"> / 89</span></b></span></td><td>47.2%</td></tr><tr><th scope="row">Nanbeige4.2-3B</th><td>4.2B reported</td><td>89<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:14.6067%"></i><b>13<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:41.5730%"></i><b>37<span class="denom"> / 89</span></b></span></td><td>41.6%</td></tr><tr><th scope="row">Gemma 4 12B QAT</th><td>12B</td><td>90<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:13.4831%"></i><b>12<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:38.2022%"></i><b>34<span class="denom"> / 89</span></b></span></td><td>38.2%</td></tr><tr><th scope="row">Qwen3.5-2B</th><td>2B</td><td>92<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:12.3596%"></i><b>11<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:31.4607%"></i><b>28<span class="denom"> / 89</span></b></span></td><td>31.5%</td></tr>-<tr class="base"><th scope="row">MiniCPM5-2B base<small>Compass v2's base</small></th><td>~2.6B</td><td>96<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:10.1124%"></i><b>9<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:23.5955%"></i><b>21<span class="denom"> / 89</span></b></span></td><td>23.6%</td></tr>+<tr class="base"><th scope="row">MiniCPM5-2B base<small>Compass Tiny v2's base</small></th><td>~2.6B</td><td>96<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:10.1124%"></i><b>9<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:23.5955%"></i><b>21<span class="denom"> / 89</span></b></span></td><td>23.6%</td></tr><tr><th scope="row">Gemma 4 E2B</th><td>4.6B</td><td>74<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:8.9888%"></i><b>8<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:50.5618%"></i><b>45<span class="denom"> / 89</span></b></span></td><td>50.6%</td></tr><tr><th scope="row">WebBrain Compass Tiny v1</th><td>2.6B</td><td>81<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:4.4944%"></i><b>4<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:47.1910%"></i><b>42<span class="denom"> / 89</span></b></span></td><td>47.2%</td></tr><tr><th scope="row">LFM2.5-2.6B base</th><td>2.6B</td><td>80<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:4.4944%"></i><b>4<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:38.2022%"></i><b>34<span class="denom"> / 89</span></b></span></td><td>38.2%</td></tr>@@ -83,12 +83,12 @@ Sorted by strict exact action, then loose match. Larger Qwen models are included<figcaption id="compass-chart-caption">100 first-turn prompts; 89 scored next-action scenarios. Strict requires the reference tool and arguments. Loose also credits name-only matches and equivalent terminal prose. Structured calls measure format, not correctness. Bars show counts, not executed browser-task success. On narrow screens, scroll the chart sideways.</figcaption></figure>-Compass v2 is one exact match behind Qwen3.8-27B and two tool-family matches ahead in these runs—a promising result at roughly one-tenth the parameter count, not a claim of general 27B-level capability.+Compass Tiny v2 is one exact match behind Qwen3.8-27B and two tool-family matches ahead in these runs—a promising result at roughly one-tenth the parameter count, not a claim of general 27B-level capability.## What's next-The BF16 research preview is private on Hugging Face. ONNX compatibility work is underway; integration is planned, not shipped. Runtime checks and training-data licensing clearance remain prerequisites.+The [BF16 research preview](https://huggingface.co/webbrain-one/webbrain-compass-tiny-v2) is private on Hugging Face. ONNX compatibility work is underway; integration is planned, not shipped. Runtime checks and training-data licensing clearance remain prerequisites.-Results combine our [earlier comparison](/blog/compact-tool-routing-models-compared), the author's separate MiniCPM5 base run, and Compass v2's [merged BF16 run and saved results](https://github.com/esokullu/webbrain/blob/main/test/llm/analysis/webbrain-compass-v2-bf16-20260910.md). Backends and precision differ; this is not a controlled speed comparison. Browser actions were not executed. Compass v2 retained one malformed-XML response and missed our earlier adapter-relative loose-match target (41 versus 43), so it remains experimental.+Results combine our [earlier comparison](/blog/compact-tool-routing-models-compared), the author's separate MiniCPM5 base run, and Compass Tiny v2's [merged BF16 run and saved results](https://github.com/esokullu/webbrain/blob/main/test/llm/analysis/webbrain-compass-v2-bf16-20260910.md). Backends and precision differ; this is not a controlled speed comparison. Browser actions were not executed. Compass Tiny v2 retained one malformed-XML response and missed our earlier adapter-relative loose-match target (41 versus 43), so it remains experimental.-Tags: #WebBrain #CompassV2 #MiniCPM5 #ToolCalling #LocalAI+Tags: #WebBrain #CompassTinyV2 #MiniCPM5 #ToolCalling #LocalAIdiff --git a/web/blog/webbrain-compass-v2/index.html b/web/blog/webbrain-compass-v2/index.htmlindex e49c2b21c..093171c33 100644--- a/web/blog/webbrain-compass-v2/index.html+++ b/web/blog/webbrain-compass-v2/index.html@@ -3,10 +3,10 @@<head><meta charset="UTF-8"><meta name="viewport" content="width=device-width, initial-scale=1.0">- <title>WebBrain Compass v2: MiniCPM5-Based Tool Routing - WebBrain Blog</title>- <meta name="description" content="Meet WebBrain Compass v2, our MiniCPM5-2B-based tool-routing model. Compare its merged BF16 results with 11 other models, including larger reference models.">- <meta name="keywords" content="WebBrain Compass v2, MiniCPM5-2B, tool calling, browser agent benchmark, local AI">- <meta property="og:title" content="Meet WebBrain Compass v2">+ <title>WebBrain Compass Tiny v2: MiniCPM5-Based Tool Routing - WebBrain Blog</title>+ <meta name="description" content="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.">+ <meta name="keywords" content="WebBrain Compass Tiny v2, MiniCPM5-2B, tool calling, browser agent benchmark, local AI">+ <meta property="og:title" content="Meet WebBrain Compass Tiny v2"><meta property="og:description" content="A small, specialized tool-routing model: 16/89 exact actions and 41/89 tool-family matches in our Compact routing benchmark."><meta property="og:type" content="article"><meta property="og:url" content="https://webbrain.one/blog/webbrain-compass-v2">@@ -17,7 +17,7 @@<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="Meet WebBrain Compass v2">+ <meta name="twitter:title" content="Meet WebBrain Compass Tiny v2"><meta name="twitter:description" content="A small, specialized tool-routing model: 16/89 exact actions and 41/89 tool-family matches in our Compact routing benchmark."><meta name="twitter:image" content="https://webbrain.one/twitter-image.png"><meta name="twitter:image:alt" content="WebBrain — Open-source AI browser agent">@@ -474,8 +474,8 @@{"@context": "https://schema.org","@type": "BlogPosting",- "headline": "WebBrain Compass v2: small model, sharper tool routing",- "description": "Meet WebBrain Compass v2, our MiniCPM5-2B-based tool-routing model. Compare its merged BF16 results with 11 other models, including larger reference models.",+ "headline": "WebBrain Compass Tiny v2: small model, sharper tool routing",+ "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",@@ -524,10 +524,10 @@<article><div class="meta">September 10, 2026 · 2 min read · <a href="/blog">← All posts</a></div>- <h1>WebBrain Compass v2: small model, sharper tool routing</h1>- <p class="lede"><strong>Meet WebBrain Compass v2</strong>, our new small model for choosing WebBrain browser tools and their arguments. Built on MiniCPM5-2B, it shows a substantial routing improvement over its base. We expect to incorporate it into WebBrain soon, subject to runtime validation and release checks.</p>+ <h1>WebBrain Compass Tiny v2: small model, sharper tool routing</h1>+ <p class="lede"><strong>Meet WebBrain Compass Tiny v2</strong>, our new small model for choosing WebBrain browser tools and their arguments. Built on MiniCPM5-2B, it shows a substantial routing improvement over its base. We expect to incorporate it into WebBrain soon, subject to runtime validation and release checks.</p><h2 id="a-focused-upgrade">A focused upgrade</h2>-<p>Compass v2 is a LoRA fine-tune of <a href="https://huggingface.co/openbmb/MiniCPM5-2B" target="_blank" rel="noopener">OpenBMB's MiniCPM5-2B</a>, approximately 2.6B parameters. We trained it for one epoch on 7,311 reviewed examples, with 213 held out for validation. Its job is specialized tool routing—not replacing a general-purpose planner.</p>+<p>Compass Tiny v2 is a LoRA fine-tune of <a href="https://huggingface.co/openbmb/MiniCPM5-2B" target="_blank" rel="noopener">OpenBMB's MiniCPM5-2B</a>, approximately 2.6B parameters. We trained it for one epoch on 7,311 reviewed examples, with 213 held out for validation. Its job is specialized tool routing—not replacing a general-purpose planner.</p><p>Against the separately tested base, the merged BF16 model moves from <strong>9 to 16 exact actions</strong> and <strong>21 to 41 tool-family matches</strong>. Structured-call coverage slips from 96 to 94 out of 100.</p><h2 id="one-comparison-all-models">One comparison, all models</h2><p>Sorted by strict exact action, then loose match. Larger Qwen models are included as reference points. Both bar columns use the same 0–89 scale.</p>@@ -559,14 +559,14 @@<thead><tr><th scope="col">Model</th><th scope="col">Parameters</th><th scope="col">Structured calls</th><th scope="col">Strict exact action</th><th scope="col">Loose tool-family match</th><th scope="col">Loose rate</th></tr></thead><tbody><tr><th scope="row">Qwen3.8-27B</th><td>27B</td><td>99<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:19.1011%"></i><b>17<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:43.8202%"></i><b>39<span class="denom"> / 89</span></b></span></td><td>43.8%</td></tr>-<tr class="compass"><th scope="row">WebBrain Compass v2<small>Merged BF16 · new</small></th><td>~2.6B</td><td>94<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:17.9775%"></i><b>16<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:46.0674%"></i><b>41<span class="denom"> / 89</span></b></span></td><td>46.1%</td></tr>+<tr class="compass"><th scope="row">WebBrain Compass Tiny v2<small>Merged BF16 · new</small></th><td>~2.6B</td><td>94<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:17.9775%"></i><b>16<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:46.0674%"></i><b>41<span class="denom"> / 89</span></b></span></td><td>46.1%</td></tr><tr><th scope="row">Qwen3.5-4B</th><td>4B</td><td>89<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:17.9775%"></i><b>16<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:40.4494%"></i><b>36<span class="denom"> / 89</span></b></span></td><td>40.4%</td></tr><tr><th scope="row">Gemma 4 E4B</th><td>7.5B</td><td>90<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:16.8539%"></i><b>15<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:48.3146%"></i><b>43<span class="denom"> / 89</span></b></span></td><td>48.3%</td></tr><tr><th scope="row">Qwen3.6-35B-A3B</th><td>35B A3B</td><td>89<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:16.8539%"></i><b>15<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:47.1910%"></i><b>42<span class="denom"> / 89</span></b></span></td><td>47.2%</td></tr><tr><th scope="row">Nanbeige4.2-3B</th><td>4.2B reported</td><td>89<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:14.6067%"></i><b>13<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:41.5730%"></i><b>37<span class="denom"> / 89</span></b></span></td><td>41.6%</td></tr><tr><th scope="row">Gemma 4 12B QAT</th><td>12B</td><td>90<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:13.4831%"></i><b>12<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:38.2022%"></i><b>34<span class="denom"> / 89</span></b></span></td><td>38.2%</td></tr><tr><th scope="row">Qwen3.5-2B</th><td>2B</td><td>92<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:12.3596%"></i><b>11<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:31.4607%"></i><b>28<span class="denom"> / 89</span></b></span></td><td>31.5%</td></tr>-<tr class="base"><th scope="row">MiniCPM5-2B base<small>Compass v2's base</small></th><td>~2.6B</td><td>96<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:10.1124%"></i><b>9<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:23.5955%"></i><b>21<span class="denom"> / 89</span></b></span></td><td>23.6%</td></tr>+<tr class="base"><th scope="row">MiniCPM5-2B base<small>Compass Tiny v2's base</small></th><td>~2.6B</td><td>96<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:10.1124%"></i><b>9<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:23.5955%"></i><b>21<span class="denom"> / 89</span></b></span></td><td>23.6%</td></tr><tr><th scope="row">Gemma 4 E2B</th><td>4.6B</td><td>74<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:8.9888%"></i><b>8<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:50.5618%"></i><b>45<span class="denom"> / 89</span></b></span></td><td>50.6%</td></tr><tr><th scope="row">WebBrain Compass Tiny v1</th><td>2.6B</td><td>81<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:4.4944%"></i><b>4<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:47.1910%"></i><b>42<span class="denom"> / 89</span></b></span></td><td>47.2%</td></tr><tr><th scope="row">LFM2.5-2.6B base</th><td>2.6B</td><td>80<span class="denom"> / 100</span></td><td><span class="score strict"><i aria-hidden="true" style="width:4.4944%"></i><b>4<span class="denom"> / 89</span></b></span></td><td><span class="score loose"><i aria-hidden="true" style="width:38.2022%"></i><b>34<span class="denom"> / 89</span></b></span></td><td>38.2%</td></tr>@@ -575,11 +575,11 @@</div><figcaption id="compass-chart-caption">100 first-turn prompts; 89 scored next-action scenarios. Strict requires the reference tool and arguments. Loose also credits name-only matches and equivalent terminal prose. Structured calls measure format, not correctness. Bars show counts, not executed browser-task success. On narrow screens, scroll the chart sideways.</figcaption></figure>-<p>Compass v2 is one exact match behind Qwen3.8-27B and two tool-family matches ahead in these runs—a promising result at roughly one-tenth the parameter count, not a claim of general 27B-level capability.</p>+<p>Compass Tiny v2 is one exact match behind Qwen3.8-27B and two tool-family matches ahead in these runs—a promising result at roughly one-tenth the parameter count, not a claim of general 27B-level capability.</p><h2 id="what-s-next">What's next</h2>-<p>The BF16 research preview is private on Hugging Face. ONNX compatibility work is underway; integration is planned, not shipped. Runtime checks and training-data licensing clearance remain prerequisites.</p>-<p>Results combine our <a href="/blog/compact-tool-routing-models-compared">earlier comparison</a>, the author's separate MiniCPM5 base run, and Compass v2's <a href="https://github.com/esokullu/webbrain/blob/main/test/llm/analysis/webbrain-compass-v2-bf16-20260910.md" target="_blank" rel="noopener">merged BF16 run and saved results</a>. Backends and precision differ; this is not a controlled speed comparison. Browser actions were not executed. Compass v2 retained one malformed-XML response and missed our earlier adapter-relative loose-match target (41 versus 43), so it remains experimental.</p>-<p>Tags: #WebBrain #CompassV2 #MiniCPM5 #ToolCalling #LocalAI</p>+<p>The <a href="https://huggingface.co/webbrain-one/webbrain-compass-tiny-v2" target="_blank" rel="noopener">BF16 research preview</a> is private on Hugging Face. ONNX compatibility work is underway; integration is planned, not shipped. Runtime checks and training-data licensing clearance remain prerequisites.</p>+<p>Results combine our <a href="/blog/compact-tool-routing-models-compared">earlier comparison</a>, the author's separate MiniCPM5 base run, and Compass Tiny v2's <a href="https://github.com/esokullu/webbrain/blob/main/test/llm/analysis/webbrain-compass-v2-bf16-20260910.md" target="_blank" rel="noopener">merged BF16 run and saved results</a>. Backends and precision differ; this is not a controlled speed comparison. Browser actions were not executed. Compass Tiny v2 retained one malformed-XML response and missed our earlier adapter-relative loose-match target (41 versus 43), so it remains experimental.</p>+<p>Tags: #WebBrain #CompassTinyV2 #MiniCPM5 #ToolCalling #LocalAI</p><div class="author-box">Written by <a href="https://emresokullu.com" target="_blank" rel="noopener">Emre Sokullu</a>. WebBrain 33.0.0 and later is GPL-3.0-or-later and open on <a href="https://github.com/webbrain-one/webbrain" target="_blank" rel="noopener">GitHub</a>.