Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

llm(free-lane): finish the gemma-4-31b-it move through the chat provider seed and llm.js

nirholas17089b935f44
diff --git a/api/_lib/llm.js b/api/_lib/llm.js
index 1db699e29..9b03e88b0 100644
--- a/api/_lib/llm.js
+++ b/api/_lib/llm.js
@@ -62,6 +62,23 @@ const GROQ_MODEL = 'qwen/qwen3.8-27b';
// Smaller model, so it sits at the END of the free section — every 70B-class
// provider gets tried before the chain steps down in capability.
const GROQ_INSTANT_MODEL = 'openai/gpt-oss-20b';
+// Third Groq rung, on a third independent per-model token pool. Groq meters
+// tokens PER MODEL at 8,000 tokens/minute, which is roughly three fact-check
+// stance calls before the rung 429s for the rest of the minute: that ceiling,
+// not any outage, is why a burst of real work fell straight past a healthy Groq
+// and down a chain whose next five rungs were dead. Each model id is a separate
+// 8k bucket, so pooling three of them triples the burst Groq absorbs before the
+// chain steps outside it. 120B rides directly behind the primary rung because
+// it is the larger model; the 20B instant lane stays at the back of the free
+// section, where a step DOWN in capability belongs.
+//
+// Measured 2026-09-04 off x-ratelimit-limit-tokens. groq/compound-mini
+// advertises 70,000 instead and looks like the obvious pick, but it is an
+// agentic model that internally routes to llama-3.3-70b-versatile, whose
+// 100,000 tokens/DAY org cap was already 99.9% spent: it 429s in production
+// while reporting a nearly-full per-minute budget. Do not add it on the
+// strength of that header.
+const GROQ_LARGE_MODEL = 'openai/gpt-oss-120b';
// Same Llama 3.3 70B on Cerebras' free tier (cloud.cerebras.ai) — optional
// rung, active when CEREBRAS_API_KEY is configured.
const CEREBRAS_MODEL = 'llama-3.3-70b';
@@ -510,6 +527,13 @@ export function providerChain({ anthropicKey, anthropicModel, grokKey = null, gr
url: 'https://api.groq.com/openai/v1/chat/completions',
model: GROQ_MODEL,
}));
+ // Same key, separate per-model token bucket. See GROQ_LARGE_MODEL.
+ chain.push(openaiCompatProvider({
+ name: 'groq#120b',
+ key: env.GROQ_API_KEY,
+ url: 'https://api.groq.com/openai/v1/chat/completions',
+ model: GROQ_LARGE_MODEL,
+ }));
}
// Same 70B family on Cerebras' free tier — a distinct quota pool from Groq,
// so one provider's daily cap doesn't take the whole 70B class down.
diff --git a/chat/src/providers.js b/chat/src/providers.js
index ba8bd72fc..62a5c1254 100644
--- a/chat/src/providers.js
+++ b/chat/src/providers.js
@@ -18,9 +18,8 @@ import {
// app now replaces this seed with the live list as soon as it loads, and falls
// back to a live id rather than to this array (see resolveDefaultModel).
export const BUILTIN_MODELS = [
- { id: 'openai/gpt-oss-20b:free', name: 'GPT OSS 20B', provider: 'Built-in', modality: 'text->text' },
- { id: 'nvidia/nemotron-3-super-120b-a12b:free', name: 'Nemotron 3 Super 120B', provider: 'Built-in', modality: 'text->text' },
{ id: 'google/gemma-4-31b-it:free', name: 'Gemma 4 31B', provider: 'Built-in', modality: 'text->text' },
+ { id: 'nvidia/nemotron-3-super-120b-a12b:free', name: 'Nemotron 3 Super 120B', provider: 'Built-in', modality: 'text->text' },
{ id: 'inclusionai/ling-3.0-flash:free', name: 'Ling 3.0 Flash', provider: 'Built-in', modality: 'text->text' },
{ id: 'nvidia/nemotron-3-ultra-550b-a55b:free', name: 'Nemotron 3 Ultra 550B', provider: 'Built-in', modality: 'text->text' },
];
diff --git a/scripts/i18n-translate.mjs b/scripts/i18n-translate.mjs
index a06c954ae..c1cfb90ce 100644
--- a/scripts/i18n-translate.mjs
+++ b/scripts/i18n-translate.mjs
@@ -220,7 +220,7 @@ const PROVIDER_DEFAULT_MODEL = {
gemini: 'gemini-2.5-flash',
vertex: 'google/gemini-2.5-flash',
groq: 'qwen/qwen3.8-27b',
- openrouter: 'openai/gpt-oss-20b:free',
+ openrouter: 'google/gemma-4-31b-it:free',
nvidia: 'meta/llama-3.3-70b-instruct',
mistral: 'mistral-small-latest',
openai: 'gpt-4o-mini',