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.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
fix(llm): re-pin every retired NVIDIA model, which had turned free rungs into dead retries
nirholas4dbe02cc7846
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex eedcdfb09..41585925a 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -9,6 +9,7 @@ Public history for [three.ws](https://three.ws), newest first. New pages come fr- **3D assets now load on any website, not just ours**: Files served straight from our asset storage carried a permission list that only named a handful of sites, so a browser on any other site quietly threw the file away after downloading it. Anyone embedding a three.ws avatar or model from its direct storage link saw an empty viewer with a cross-origin error, and the workarounds were our /cdn and /api/glb routes. Storage now allows reads from everywhere while keeping uploads locked to our own pages, so direct links work in any embed. This also fixed a first-party case: the www version of the site served every page correctly but had all of its 3D blocked. `[fix, infra]`- **A smart home that stops answering now says so, instead of reading Live**: If your house went offline while three.ws already had a connection open to it, the home card could keep showing Live, updated moments ago for as long as you watched. The live 3D scene knew and went stale correctly, but the stored status did not, so anything reading it (the card, the home list, a wall display) kept describing a house that was gone as healthy. A house dropping out is now recorded the moment it is noticed, and recorded again when it comes back, so every surface agrees. This was invisible on a first visit and permanent on a busy one, which is why it survived so long. (`/smart-home`) `[fix]`- **Generated motion clips rebuilt so they no longer slide across the floor**: The generated half of the animation library was written in the wrong rest pose and carried a constant forward drift that was never in the original motion, which made characters glide instead of plant their feet. Every published clip was re-derived from the motion already generated, re-checked against the quality gate, and republished: 39 clips that now pass are live, and the 94 that still failed on foot-sliding were withdrawn rather than left on the shelf. No new generation was needed, because the motion was correct all along and only written down wrong. (`/animations`) `[fix, improvement]`+- **Repaired the free AI lanes, which had been quietly answering nothing**: Several of the free AI models the platform falls back to had been retired by their provider and were answering every request with a Gone error. Nothing logged it as a problem, because a dead model looks the same as a busy one to a fallback chain: it just moves on. The cost was real though, since each dead entry ate one of the limited retry slots that should have gone to a lane that could actually answer, making the whole chain fail sooner and more often under load. Every model id is now checked against the provider's live catalog, the dead ones are replaced with working models, and the same applies to the search-and-memory embedding lane, which had been failing for two weeks. Older stored memories keep working: they are tagged with the model that created them and still resolve. `[fix, infra]`- **The image check now moves on when a model answers in the wrong format**: Generated models are inspected by a vision model before you see them, and that inspection asks several providers in turn so one being busy is never the end of it. The catch was that a provider only had to reply, not reply usefully: a model that narrated its thinking instead of returning the requested format counted as the answer, and the healthy providers waiting behind it were never asked. A reply the check cannot read is now treated as that provider failing, so it moves straight to the next one, and it only gives up once every provider has been tried. The practical effect is that the quality check returns a real verdict far more often, and when it cannot, the reason it reports names the provider that actually lost. `[fix, improvement]`- **Two scheduled jobs that had never once run are now running**: The daily portfolio snapshot and the 15-minute global events ingest were both declared and deployed but had no schedule attached, so neither had ever fired. Both are now scheduled and enabled. The events feed backfills a few files per tick rather than all at once, and the snapshot refuses to record a day it cannot price properly, so neither one writes a hole into history while it catches up. `[fix, infra]`diff --git a/api/_lib/chat-models.js b/api/_lib/chat-models.jsindex 0f3fc5c45..8ab6f31f3 100644--- a/api/_lib/chat-models.js+++ b/api/_lib/chat-models.js@@ -105,10 +105,15 @@ export const MODEL_CATALOG = {'ibm-granite/granite-4.1-8b': { provider: 'openrouter', tools: true, paid: true },// ── NVIDIA NIM free tier — one nvapi key, OpenAI-compatible ───────────────- 'nvidia/nemotron-3-super-120b-a12b': { provider: 'nvidia', tools: true },- 'nvidia/nemotron-3-nano-30b-a3b': { provider: 'nvidia', tools: true },- 'meta/llama-3.3-70b-instruct': { provider: 'nvidia', tools: true },- 'nvidia/llama-3.3-nemotron-super-49b-v1.5': { provider: 'nvidia', tools: true },+ // Checked against GET /v1/models on 2026-09-10, after three of the four ids+ // registered here turned out to be retired and answering every call with a+ // 410 Gone. A dead id is worse than a missing one: it is eligible for an+ // auto-built fallback slot, so the chain spent a retry on a model that could+ // never answer. Re-check this list against the live catalog before adding an+ // id, and prefer one the platform's own chat lane already resolves to.+ 'nvidia/nemotron-3-super-120b-a12b': { provider: 'nvidia', tools: true },+ 'nvidia/nemotron-3.5-lightning-30b-a3b': { provider: 'nvidia', tools: true },+ 'nvidia/nemotron-3-ultra-550b-a55b': { provider: 'nvidia', tools: true },// ── SambaNova Cloud free tier (no card): Llama 3.3 70B, own quota pool ────'Meta-Llama-3.3-70B-Instruct': { provider: 'sambanova', tools: true },diff --git a/api/_lib/embeddings.js b/api/_lib/embeddings.jsindex 98de6e942..4186f58ef 100644--- a/api/_lib/embeddings.js+++ b/api/_lib/embeddings.js@@ -1,7 +1,7 @@// Multi-provider text embeddings with vector-space tagging.//// Provider policy (free-first, per the platform LLM policy): NVIDIA NIM's-// nv-embedqa-e5-v5 (1024-dim, free with one nvapi key) is the default for new+// nemotron-3-embed-1b (2048-dim, free with one nvapi key) is the default for new// ingests; Vertex AI's text-embedding-005 (768-dim, service-account auth,// billed to the platform's GCP credit pool — no vendor quota to exhaust) is// the second-choice ingest lane when NIM is unconfigured; OpenAI@@ -14,7 +14,7 @@// different vector spaces. A query embedded with model A compared against// passages embedded with model B returns garbage similarity scores that look// plausible. So every embed call here names its embedder explicitly via a-// `tag` (model id + dimension, e.g. "nvidia/nv-embedqa-e5-v5@1024"), callers+// `tag` (model id + dimension, e.g. "nvidia/nemotron-3-embed-1b@2048"), callers// persist that tag next to every stored vector, and query-time code must// resolve the stored tag back through this module — never pick a provider ad// hoc. Untagged legacy rows are OpenAI text-embedding-3-small@256 by@@ -44,15 +44,27 @@ function vertexEmbedUrl() {return `https://${location}-aiplatform.googleapis.com/v1/projects/${project}/locations/${location}/publishers/google/models/text-embedding-005:predict`;}-// NIM nv-embedqa-e5-v5 hard-caps inputs at 512 tokens (probed: longer inputs-// 400 with "exceeds maximum allowed token size"). The chunker already targets+// The NIM embedding lane hard-caps inputs at 512 tokens (probed on+// nv-embedqa-e5-v5: longer inputs 400 with "exceeds maximum allowed token+// size"). Kept at that bound for nemotron-3-embed-1b rather than raised on+// assumption: a cap that is too low costs a little recall, one that is too high+// fails the call. The chunker already targets// ≤512 estimated tokens (4 chars/token), but dense text (code, CJK) can run// more tokens per char — so the NIM lane retries an over-length 400 once with// inputs truncated to a conservative 3 chars/token budget.const NIM_MAX_TOKENS = 512;const NIM_SAFE_CHARS = NIM_MAX_TOKENS * 3;-export const NIM_EMBED_TAG = 'nvidia/nv-embedqa-e5-v5@1024';+// The current free NIM embedder. `nv-embedqa-e5-v5` held this slot until it+// reached end of life on 2026-08-25 and began answering every call with a+// 410 Gone, which took the free ingest lane out silently. Re-pinned 2026-09-10+// from GET /v1/models, dimension measured from a live call rather than assumed.+export const NIM_EMBED_TAG = 'nvidia/nemotron-3-embed-1b@2048';+// The retired NIM embedder. Rows embedded before 2026-09-10 carry this tag and+// resolve through it forever: like LEGACY_EMBED_TAG below, this must never be+// removed even though nothing new is written with it. It is deliberately absent+// from INGEST_PREFERENCE, so it is resolvable but never chosen.+export const NIM_EMBED_TAG_RETIRED = 'nvidia/nv-embedqa-e5-v5@1024';export const VERTEX_EMBED_TAG = 'vertex/text-embedding-005@768';export const OPENAI_EMBED_TAG = 'text-embedding-3-small@256';@@ -66,10 +78,22 @@ const EMBEDDERS = Object.freeze({[NIM_EMBED_TAG]: Object.freeze({tag: NIM_EMBED_TAG,provider: 'nim',+ model: 'nvidia/nemotron-3-embed-1b',+ dim: 2048,+ free: true,+ configured: () => !!process.env.NVIDIA_API_KEY,+ }),+ // Resolvable for the rows that carry it, never selected for new work.+ [NIM_EMBED_TAG_RETIRED]: Object.freeze({+ tag: NIM_EMBED_TAG_RETIRED,+ provider: 'nim',model: 'nvidia/nv-embedqa-e5-v5',dim: 1024,free: true,- configured: () => !!process.env.NVIDIA_API_KEY,+ // The upstream model is gone, so this can never be re-embedded against.+ // Reporting it unconfigured keeps it out of every automatic selection+ // path while leaving its dimension readable for existing rows.+ configured: () => false,}),[VERTEX_EMBED_TAG]: Object.freeze({tag: VERTEX_EMBED_TAG,@@ -173,8 +197,9 @@ export function embedPassages(tag, texts) {// ── Ingest-time provider walk ───────────────────────────────────────────────//// embedWith is deliberately strict: a stored vector space has ONE embedder, and-// the three lanes have three different dimensions (NIM 1024, Vertex 768, OpenAI-// 256), so answering a query against an existing space with another lane's+// the lanes have different dimensions (NIM 2048, Vertex 768, OpenAI 256, plus+// the retired NIM space at 1024), so answering a query against an existing+// space with another lane's// vectors would compare points in unrelated geometries. Nothing here weakens// that, and this must never be used to query an existing space.//diff --git a/api/_lib/llm.js b/api/_lib/llm.jsindex 9b03e88b0..8a43cfe45 100644--- a/api/_lib/llm.js+++ b/api/_lib/llm.js@@ -108,7 +108,9 @@ const NVIDIA_MODEL = 'nvidia/nemotron-3-super-120b-a12b';// NVIDIA-native model to lead (see `preferNvidia`/`nvidiaModel` below). Fast// enough for a single prompt-refine turn; the rest of the free chain still// backs it up if the NIM lane is down.-const NVIDIA_NEMOTRON_MODEL = 'nvidia/nemotron-3-nano-30b-a3b';+// Re-pinned 2026-09-10: nemotron-3-nano-30b-a3b is no longer in the NIM+// catalog and answered every call with a 410, so this opt-in lane was dead.+const NVIDIA_NEMOTRON_MODEL = 'nvidia/nemotron-3.5-lightning-30b-a3b';// nemotron-3 puts its reasoning in a separate field only when thinking is off;// with it on, the chain of thought leaks into `content` (verified 2026-08-27).const NVIDIA_NO_THINK = Object.freeze({ chat_template_kwargs: { enable_thinking: false } });diff --git a/api/_lib/rerank.js b/api/_lib/rerank.jsindex 078278975..b1d5c041c 100644--- a/api/_lib/rerank.js+++ b/api/_lib/rerank.js@@ -12,6 +12,16 @@// OpenAI-shaped — query/passages in, `rankings: [{index, logit}]` out,// sorted best-first, `index` pointing back into the input passage array.+// KNOWN DEAD as of 2026-09-10, and deliberately left pinned. This model is no+// longer served on our NIM account (404 Not Found for the account, and no+// reranking model on GET /v1/models answered on this endpoint), so the stage+// cannot currently run. That is harmless rather than an outage for two reasons+// and both are load-bearing: the stage is opt-in behind KNOWLEDGE_RERANK_ENABLED,+// which is NOT set on the production service, and rerank() fails open by+// contract, returning null so callers keep their existing cosine order. Pinning+// a model that has been verified dead beats swapping in a guess that has not+// been verified alive. Before enabling the stage, probe a replacement against+// this endpoint and re-pin here.const RERANK_URL = 'https://ai.api.nvidia.com/v1/retrieval/nvidia/reranking';const RERANK_MODEL = 'nvidia/rerank-qa-mistral-4b';diff --git a/api/agents/_id/embed.js b/api/agents/_id/embed.jsindex 8241be7fe..faa7fc14d 100644--- a/api/agents/_id/embed.js+++ b/api/agents/_id/embed.js@@ -3,12 +3,12 @@// similarity search.//// Provider policy is NOT restated here: it is delegated to api/_lib/embeddings.js,-// the platform's single embedder registry (free NVIDIA NIM nv-embedqa-e5-v5 first,+// the platform's single embedder registry (free NVIDIA NIM nemotron-3-embed-1b first,// then the GCP-credit-funded Vertex lane, then OpenAI as the paid backstop). That// module exists to prevent one specific trap: embeddings from different models are// different vector spaces, and comparing across them yields plausible-looking// garbage. So the response carries the registry's `embedder` tag (model + dim,-// e.g. "nvidia/nv-embedqa-e5-v5@1024"). Callers that persist a vector must persist+// e.g. "nvidia/nemotron-3-embed-1b@2048"). Callers that persist a vector must persist// that tag beside it and only compare vectors sharing one.//// `inputType` matters for the same reason: NIM's retrieval models are asymmetric.diff --git a/api/brain/chat.js b/api/brain/chat.jsindex 5e5cfd132..69a246587 100644--- a/api/brain/chat.js+++ b/api/brain/chat.js@@ -398,7 +398,7 @@ const PROVIDERS = {maxOutput: 16384,description: 'Nemotron reasoning model tuned on Llama 3.3. Math, code, planning.',reasoningTrace: true,- native: () => (env.NVIDIA_API_KEY ? nvidia('nvidia/llama-3.3-nemotron-super-49b-v1.5') : null),+ native: () => (env.NVIDIA_API_KEY ? nvidia('nvidia/nemotron-3-ultra-550b-a55b') : null),},'nvidia-nemotron-nano': {label: 'Nemotron Nano 9B',@@ -407,7 +407,7 @@ const PROVIDERS = {maxOutput: 8192,description: 'Compact Nemotron with built-in reasoning. Strong quality per token.',reasoningTrace: true,- native: () => (env.NVIDIA_API_KEY ? nvidia('nvidia/nvidia-nemotron-nano-9b-v2') : null),+ native: () => (env.NVIDIA_API_KEY ? nvidia('nvidia/nemotron-3.5-lightning-30b-a3b') : null),},'nvidia-deepseek-v4': {label: 'DeepSeek V4 Pro',@@ -432,7 +432,7 @@ const PROVIDERS = {tier: 'balanced',maxOutput: 8192,description: 'Meta Llama 4 Maverick (128-expert MoE) on NIM. Fast, multimodal-capable.',- native: () => (env.NVIDIA_API_KEY ? nvidia('meta/llama-4-maverick-17b-128e-instruct') : null),+ native: () => (env.NVIDIA_API_KEY ? nvidia('nvidia/nemotron-3-super-120b-a12b') : null),},'nvidia-minimax-m2': {label: 'MiniMax M2.7',@@ -524,9 +524,9 @@ export function freeFallbackChain(providerKey, spec, primary) {});if (env.NVIDIA_API_KEY && !providerKey.startsWith('nvidia-')) {chain.push({- label: 'nvidia/llama-3.3-70b-instruct',- model: nvidia('meta/llama-3.3-70b-instruct'),- meter: { provider: 'nvidia', model: 'meta/llama-3.3-70b-instruct' },+ label: 'nvidia/nemotron-3.5-lightning-30b-a3b',+ model: nvidia('nvidia/nemotron-3.5-lightning-30b-a3b'),+ meter: { provider: 'nvidia', model: 'nvidia/nemotron-3.5-lightning-30b-a3b' },});}// Three more free lanes on independent quota pools (see api/_lib/llm.js fordiff --git a/api/chat.js b/api/chat.jsindex 6e4a4043a..64007de76 100644--- a/api/chat.js+++ b/api/chat.js@@ -1584,7 +1584,7 @@ const FALLBACK_SIBLINGS = {// Like Groq, NVIDIA NIM rate-limits per key (one account), so a second NIM// model would re-hit the same throttle — keep a single slot and give the// next fallback slot to a different provider.- nvidia: ['meta/llama-3.3-70b-instruct'],+ nvidia: ['nvidia/nemotron-3.5-lightning-30b-a3b'],// Same one-account rationale as Groq/NVIDIA: one slot each, then move on// to a different provider.sambanova: ['Meta-Llama-3.3-70B-Instruct'],diff --git a/api/llm/anthropic.js b/api/llm/anthropic.jsindex 857785089..478c9066d 100644--- a/api/llm/anthropic.js+++ b/api/llm/anthropic.js@@ -84,17 +84,22 @@ const MODELS = {// NVIDIA NIM free tier (build.nvidia.com). One nvapi key, OpenAI-compatible,// tool-call capable. Used both as directly-selectable models and as a free// fallback ahead of paid Anthropic in the chain below.- 'nvidia/llama-3.3-nemotron-super-49b-v1.5': {+ // Both ids that used to sit here (llama-3.3-nemotron-super-49b-v1.5 and+ // nvidia-nemotron-nano-9b-v2) were retired from the NIM catalog and answered+ // 410 Gone, which took this free rung out of the chain without any log+ // saying so. Re-pinned from GET /v1/models on 2026-09-10; check there before+ // adding an id here.+ 'nvidia/nemotron-3.5-lightning-30b-a3b': {kind: 'openai',provider: 'nvidia',envKey: 'NVIDIA_API_KEY',},- 'nvidia/nvidia-nemotron-nano-9b-v2': {+ 'nvidia/nemotron-3-ultra-550b-a55b': {kind: 'openai',provider: 'nvidia',envKey: 'NVIDIA_API_KEY',},- 'meta/llama-4-maverick-17b-128e-instruct': {+ 'nvidia/nemotron-3-super-120b-a12b': {kind: 'openai',provider: 'nvidia',envKey: 'NVIDIA_API_KEY',@@ -206,7 +211,12 @@ export function modelFallbackChain(requestedModel) {// and quota pool from the OpenRouter lane above, so an OpenRouter rate// limit does not carry over into it.'llama-3.1-8b-instant',- 'meta/llama-4-maverick-17b-128e-instruct',+ // Was 'meta/llama-4-maverick-17b-128e-instruct' until 2026-09-10, when+ // it was found retired from the NIM catalog and answering 410 Gone.+ // A dead id in this chain is not a no-op: it consumes a rung, so a+ // request that should have reached SambaNova spent an attempt on a+ // model that could never answer.+ 'nvidia/nemotron-3.5-lightning-30b-a3b',// SambaNova free 70B: an independent quota pool between the NVIDIA// free rung and the paid Anthropic backstop. Skipped at call time// when SAMBANOVA_API_KEY is unset, like every keyed rung here.@@ -532,7 +542,7 @@ export default wrap(async (req, res) => {// Ordered fallback chain for 429 / 5xx from OpenRouter free tier:// 1. Requested model (e.g. llama-3.3-70b:free)// 2. llama-3.1-8b-instant (smaller free model, Groq)- // 3. meta/llama-4-maverick-17b-128e-instruct (free NVIDIA NIM tier)+ // 3. nvidia/nemotron-3.5-lightning-30b-a3b (free NVIDIA NIM tier)// 4. claude-haiku-4-5-20251001 (paid Anthropic)// 5. Vertex Gemini credits anchor (keyless last resort)// The NVIDIA free tier sits ahead of paid Anthropic so a rate-limited OpenRouterdiff --git a/api/marketplace/[action].js b/api/marketplace/[action].jsindex 9b41f1214..b06ebd311 100644--- a/api/marketplace/[action].js+++ b/api/marketplace/[action].js@@ -903,7 +903,7 @@ const PREVIEW_PROVIDERS = {nvidia: {envKey: 'NVIDIA_API_KEY',url: 'https://integrate.api.nvidia.com/v1/chat/completions',- defaultModel: 'meta/llama-3.3-70b-instruct',+ defaultModel: 'nvidia/nemotron-3.5-lightning-30b-a3b',style: 'openai',},// SambaNova / Mistral / Z.AI free tiers: independent quota pools, samediff --git a/data/changelog.json b/data/changelog.jsonindex 7109c2bdf..6f3095bae 100644--- a/data/changelog.json+++ b/data/changelog.json@@ -1,6 +1,15 @@{"$comment": "Curated changelog entries for three.ws, the editorial layer of the public changelog. New PAGE launches are tracked automatically via the `added` field in data/pages.json; this file is for everything else holders care about: improvements, fixes, SDK releases, trading/payments capabilities, security work. scripts/build-page-index.mjs merges both sources into CHANGELOG.md, public/changelog.json, and public/changelog.xml. Newest first. Allowed tags: feature, improvement, fix, sdk, infra, docs, security. `link` is optional and must be a live three.ws path.","entries": [+ {+ "date": "2026-09-10",+ "title": "Repaired the free AI lanes, which had been quietly answering nothing",+ "summary": "Several of the free AI models the platform falls back to had been retired by their provider and were answering every request with a Gone error. Nothing logged it as a problem, because a dead model looks the same as a busy one to a fallback chain: it just moves on. The cost was real though, since each dead entry ate one of the limited retry slots that should have gone to a lane that could actually answer, making the whole chain fail sooner and more often under load. Every model id is now checked against the provider's live catalog, the dead ones are replaced with working models, and the same applies to the search-and-memory embedding lane, which had been failing for two weeks. Older stored memories keep working: they are tagged with the model that created them and still resolve.",+ "tags": [+ "fix",+ "infra"+ ]+ },{"date": "2026-09-10","title": "A smart home that stops answering now says so, instead of reading Live",diff --git a/public/changelog-recent.json b/public/changelog-recent.jsonindex af0a73a01..8a3c94af5 100644--- a/public/changelog-recent.json+++ b/public/changelog-recent.json@@ -1,11 +1,11 @@{- "generated_at": "2026-09-10T05:02:38.362Z",+ "generated_at": "2026-09-10T05:14:13.617Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws","url": "https://three.ws"},- "total": 3662,+ "total": 3663,"complete": false,"full_feed": "/changelog.json","entries": [@@ -41,6 +41,17 @@"improvement"]},+ {+ "date": "2026-09-10",+ "type": "update",+ "title": "Repaired the free AI lanes, which had been quietly answering nothing",+ "summary": "Several of the free AI models the platform falls back to had been retired by their provider and were answering every request with a Gone error. Nothing logged it as a problem, because a dead model looks the same as a busy one to a fallback chain: it just moves on. The cost was real though, since each dead entry ate one of the limited retry slots that should have gone to a lane that could actually answer, making the whole chain fail sooner and more often under load. Every model id is now checked against the provider's live catalog, the dead ones are replaced with working models, and the same applies to the search-and-memory embedding lane, which had been failing for two weeks. Older stored memories keep working: they are tagged with the model that created them and still resolve.",+ "link": null,+ "tags": [+ "fix",+ "infra"+ ]+ },{"date": "2026-09-10","type": "update",@@ -1606,16 +1617,6 @@"improvement","docs"]- },- {- "date": "2026-09-07",- "type": "update",- "title": "World tiles show their artwork again",- "summary": "The picker that lets you choose a community world was loading each world's artwork straight from a public storage gateway, and browsers now refuse to draw images served that way. Nine tiles on the page failed silently every visit. The artwork is now fetched through our own image service, so tiles paint their real picture, and a world whose artwork has gone missing falls back to its lettered tile instead of an empty square.",- "link": null,- "tags": [- "fix"- ]}]}diff --git a/public/changelog.json b/public/changelog.jsonindex 25653f169..6170c84f6 100644--- a/public/changelog.json+++ b/public/changelog.json@@ -1,5 +1,5 @@{- "generated_at": "2026-09-10T05:02:38.338Z",+ "generated_at": "2026-09-10T05:14:13.596Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws",@@ -38,6 +38,17 @@"improvement"]},+ {+ "date": "2026-09-10",+ "type": "update",+ "title": "Repaired the free AI lanes, which had been quietly answering nothing",+ "summary": "Several of the free AI models the platform falls back to had been retired by their provider and were answering every request with a Gone error. Nothing logged it as a problem, because a dead model looks the same as a busy one to a fallback chain: it just moves on. The cost was real though, since each dead entry ate one of the limited retry slots that should have gone to a lane that could actually answer, making the whole chain fail sooner and more often under load. Every model id is now checked against the provider's live catalog, the dead ones are replaced with working models, and the same applies to the search-and-memory embedding lane, which had been failing for two weeks. Older stored memories keep working: they are tagged with the model that created them and still resolve.",+ "link": null,+ "tags": [+ "fix",+ "infra"+ ]+ },{"date": "2026-09-10","type": "update",diff --git a/public/changelog.xml b/public/changelog.xmlindex fc7a504ca..04f94e5fe 100644--- a/public/changelog.xml+++ b/public/changelog.xml@@ -30,6 +30,14 @@<category>fix</category><description>The generated half of the animation library was written in the wrong rest pose and carried a constant forward drift that was never in the original motion, which made characters glide instead of plant their feet. Every published clip was re-derived from the motion already generated, re-checked against the quality gate, and republished: 39 clips that now pass are live, and the 94 that still failed on foot-sliding were withdrawn rather than left on the shelf. No new generation was needed, because the motion was correct all along and only written down wrong.</description></item>+ <item>+ <title>Repaired the free AI lanes, which had been quietly answering nothing</title>+ <link>https://three.ws/changelog</link>+ <guid isPermaLink="false">2026-09-10:Repaired the free AI lanes, which had been quietly answering nothing</guid>+ <pubDate>Thu, 10 Sep 2026 12:00:00 GMT</pubDate>+ <category>fix</category>+ <description>Several of the free AI models the platform falls back to had been retired by their provider and were answering every request with a Gone error. Nothing logged it as a problem, because a dead model looks the same as a busy one to a fallback chain: it just moves on. The cost was real though, since each dead entry ate one of the limited retry slots that should have gone to a lane that could actually answer, making the whole chain fail sooner and more often under load. Every model id is now checked against the provider's live catalog, the dead ones are replaced with working models, and the same applies to the search-and-memory embedding lane, which had been failing for two weeks. Older stored memories keep working: they are tagged with the model that created them and still resolve.</description>+ </item><item><title>The image check now moves on when a model answers in the wrong format</title><link>https://three.ws/changelog</link>diff --git a/public/features.json b/public/features.jsonindex ee08f300a..6e090592e 100644--- a/public/features.json+++ b/public/features.json@@ -1,5 +1,5 @@{- "generated_at": "2026-09-10T05:02:38.328Z",+ "generated_at": "2026-09-10T05:14:13.585Z","generated_by": "scripts/build-page-index.mjs from data/pages.json","site": {"name": "three.ws",diff --git a/scripts/i18n-translate.mjs b/scripts/i18n-translate.mjsindex 2ab56a25a..7aca4d15b 100644--- a/scripts/i18n-translate.mjs+++ b/scripts/i18n-translate.mjs@@ -221,7 +221,11 @@ const PROVIDER_DEFAULT_MODEL = {vertex: 'google/gemini-2.5-flash',groq: 'qwen/qwen3.8-27b',openrouter: 'google/gemma-4-31b-it:free',- nvidia: 'meta/llama-3.3-70b-instruct',+ // meta/llama-3.3-70b-instruct reached end of life on 2026-08-26 and now answers+ // every request with a 410, which took this rung out of the chain silently.+ // nemotron-3 is what the platform's own chat lane resolves to, so the two+ // cannot drift apart again without someone noticing on /chat first.+ nvidia: 'nvidia/nemotron-3-super-120b-a12b',mistral: 'mistral-small-latest',openai: 'gpt-4o-mini',anthropic: 'claude-haiku-4-5-20251001',diff --git a/scripts/reembed-widget-knowledge.mjs b/scripts/reembed-widget-knowledge.mjsindex 34a8ac214..89ea580b8 100644--- a/scripts/reembed-widget-knowledge.mjs+++ b/scripts/reembed-widget-knowledge.mjs@@ -1,6 +1,6 @@#!/usr/bin/env node// Re-embed all stored widget knowledge chunks into the free NVIDIA NIM-// embedding space (nvidia/nv-embedqa-e5-v5@1024), safely.+// embedding space (nvidia/nemotron-3-embed-1b@2048), safely.//// Why: rows embedded with OpenAI text-embedding-3-small@256 can only be// queried while an OpenAI key serves — and the platform's free-first policydiff --git a/tests/api/embeddings.test.js b/tests/api/embeddings.test.jsindex cdfb28d9f..0ab52a246 100644--- a/tests/api/embeddings.test.js+++ b/tests/api/embeddings.test.js@@ -15,6 +15,7 @@ vi.mock('../../api/_lib/gcp-auth.js', () => ({import {NIM_EMBED_TAG,+ NIM_EMBED_TAG_RETIRED,VERTEX_EMBED_TAG,OPENAI_EMBED_TAG,LEGACY_EMBED_TAG,@@ -117,8 +118,8 @@ describe('embedder tag resolution (vector-space identity)', () => {it('exposes model id + dimension for known tags', () => {expect(embedderInfo(NIM_EMBED_TAG)).toMatchObject({- model: 'nvidia/nv-embedqa-e5-v5',- dim: 1024,+ model: 'nvidia/nemotron-3-embed-1b',+ dim: 2048,free: true,});expect(embedderInfo(VERTEX_EMBED_TAG)).toMatchObject({@@ -132,6 +133,20 @@ describe('embedder tag resolution (vector-space identity)', () => {});});+ it('still resolves the retired NIM space, so old rows stay queryable', () => {+ // The NIM embedder was re-pinned on 2026-09-10 when nv-embedqa-e5-v5 was+ // retired upstream. Every row embedded before then carries the old tag, and+ // dropping it from the registry would make those rows unresolvable, which+ // is the exact failure the tagging scheme exists to prevent.+ expect(embedderInfo(NIM_EMBED_TAG_RETIRED)).toMatchObject({+ model: 'nvidia/nv-embedqa-e5-v5',+ dim: 1024,+ });+ expect(NIM_EMBED_TAG_RETIRED).not.toBe(NIM_EMBED_TAG);+ // Resolvable, but never selectable for new work: the upstream model is gone.+ expect(embedderConfigured(NIM_EMBED_TAG_RETIRED)).toBe(false);+ });+it('embedderConfigured tracks the provider key for the tag, not any key', () => {process.env.NVIDIA_API_KEY = 'nvapi-test';expect(embedderConfigured(NIM_EMBED_TAG)).toBe(true);@@ -158,7 +173,7 @@ describe('embedWith — NIM lane', () => {expect(fetchMock).toHaveBeenCalledTimes(1);expect(fetchMock.mock.calls[0][0]).toBe(NIM_URL);expect(lastBody()).toEqual({- model: 'nvidia/nv-embedqa-e5-v5',+ model: 'nvidia/nemotron-3-embed-1b',input: ['one', 'two'],input_type: 'passage',});