Repository navigation
fix: retarget eco SIMPLE's second free rung off the delisted nemotron-3-nano-30b - #3
Merged
Merged
Conversation
…-3-nano-30b blockrun delisted nvidia/nemotron-3-nano-30b on 2026-09-08: NVIDIA deprovisioned it for the account (a structured per-account 404), nine days after it was listed as the fastest free model. It was still ecoTiers.SIMPLE.fallback[0]. Nothing looked broken because the gateway redirects the id — which is the shape that silently defeats a host's /exclude. ecoTiers.SIMPLE.fallback[0] nemotron-3-nano-30b -> nemotron-3-nano-omni-30b-a3b-reasoning nano-omni is the gateway's own redirect target for nano-30b (both redirect tables, the retired entries' fallbackModel, and the eco.SIMPLE row of the savings mix), so the router and the gateway keep naming the same model — and the model actually served on this rung does not change. model-capabilities.ts: drops nano-30b (no longer in the catalog) and carries a supportsVision:false override for nano-omni. The catalog tags it vision, but ClawRouter's 2026-08-31 64x64 probe got it right 1 of 4 on Base and had the image silently dropped on Solana. Without the override this swap would have started routing eco SIMPLE image turns to it (the regenerated snapshot showed two vision prompts moving off glm-5.3-flash). With it, the decision snapshot moves by exactly one candidate id. Hand-applied rather than regenerated: a full sync against today's catalog rewrites ~80 unrelated lines. The two entries match the generator's output byte for byte. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
VickyXAI
added a commit
to BlockRunAI/blockrun-llm
that referenced
this pull request
Sep 29, 2026
… live free lineup (#73) blockrun delisted nvidia/nemotron-3-nano-30b on 2026-09-08 (NVIDIA deprovisioned it for the account). The gateway redirects the id to nano-omni. Code - router_adapter.FREE_TIERS: nano-30b was SIMPLE's primary and a fallback in MEDIUM, COMPLEX and REASONING. The adapter already dropped it at runtime, since it is not in the catalog, so SIMPLE had silently been opening on lightning. Every slot now goes to nemotron-3-nano-omni-30b-a3b-reasoning: the gateway's own redirect target, the same family and size, and the model those slots were actually served by. The 2026-08-31 reason for excluding it (it answered as nano-30b) is gone with nano-30b. - router_core: re-synced to router-core e38958b (BlockRunAI/router-core#3). eco SIMPLE fallback[0] goes nano-30b -> nano-omni; nano-30b is dropped from model_capabilities; nano-omni gets the supportsVision:false override. The decision fixture is a verbatim copy of upstream's regenerated snapshot, and the parity test passes on it. - New test: FREE_TIERS may only name ids from the $0 chat set /v1/models published on 2026-09-29, and never nano-30b. The existing membership check compared the table against a hand-kept list and could not catch this. Docs - README "Available free models", the `free` routing-profile row, the hero line, both code examples, the NVIDIA section and the FAQ count now name the six live free models instead of the 08-12 lineup (step-3.7-flash, mistral-nemotron, nemotron-nano v2, gpt-oss). gpt-oss-120b/20b were retired 2026-09-03, so their "direct calls still work" rows and privacy note are gone. Counts use the models.free marker. - client.py streaming docstring: example ids moved off the retired deepseek-v4-flash / llama-4-maverick. Co-authored-by: 1bcMax <viewitter@gmail.com> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
VickyXAI
added a commit
to BlockRunAI/blockrun-llm
that referenced
this pull request
Sep 30, 2026
… live free lineup (#73) blockrun delisted nvidia/nemotron-3-nano-30b on 2026-09-08 (NVIDIA deprovisioned it for the account). The gateway redirects the id to nano-omni. Code - router_adapter.FREE_TIERS: nano-30b was SIMPLE's primary and a fallback in MEDIUM, COMPLEX and REASONING. The adapter already dropped it at runtime, since it is not in the catalog, so SIMPLE had silently been opening on lightning. Every slot now goes to nemotron-3-nano-omni-30b-a3b-reasoning: the gateway's own redirect target, the same family and size, and the model those slots were actually served by. The 2026-08-31 reason for excluding it (it answered as nano-30b) is gone with nano-30b. - router_core: re-synced to router-core e38958b (BlockRunAI/router-core#3). eco SIMPLE fallback[0] goes nano-30b -> nano-omni; nano-30b is dropped from model_capabilities; nano-omni gets the supportsVision:false override. The decision fixture is a verbatim copy of upstream's regenerated snapshot, and the parity test passes on it. - New test: FREE_TIERS may only name ids from the $0 chat set /v1/models published on 2026-09-29, and never nano-30b. The existing membership check compared the table against a hand-kept list and could not catch this. Docs - README "Available free models", the `free` routing-profile row, the hero line, both code examples, the NVIDIA section and the FAQ count now name the six live free models instead of the 08-12 lineup (step-3.7-flash, mistral-nemotron, nemotron-nano v2, gpt-oss). gpt-oss-120b/20b were retired 2026-09-03, so their "direct calls still work" rows and privacy note are gone. Counts use the models.free marker. - client.py streaming docstring: example ids moved off the retired deepseek-v4-flash / llama-4-maverick. Co-authored-by: 1bcMax <viewitter@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
blockrun delisted
nvidia/nemotron-3-nano-30bon 2026-09-08 (NVIDIA deprovisioned it for the account: a structured per-account 404). It was stillecoTiers.SIMPLE.fallback[0]. The gateway redirects the id, so nothing errored, and that is how a host's/excludegets silently defeated.Changes
config.tsecoTiers.SIMPLE.fallback[0]=nvidia/nemotron-3-nano-30bnvidia/nemotron-3-nano-omni-30b-a3b-reasoningfallbackModel, and theeco.SIMPLErow ofsavings-mix.json). It follows the head's rule that the router and the gateway name the same model. The model served on this rung does not change.model-capabilities.tsnvidia/nemotron-3-nano-30bentry/v1/models.model-capabilities.ts+scripts/capability-overrides.jsonsupportsVision: true(from catalog categories)supportsVision: falseoverride, with reasonglm-5.3-flash.decisions.snapshot.jsondist/I hand-applied the capability entries instead of running
sync-model-capabilities.mjs. A full sync against today's catalog rewrites about 80 unrelated lines. The two entries match the generator's output byte for byte (I checked by running it and diffing).Checks
npm run check(typecheck, 104/104 tests, build): passgit diff --exit-code -- distafter the build: cleanConsumers
ClawRouter's PR (
chore/savings-and-free-lineup) re-pins@blockrun/router-coreto this branch's heade38958b. Itsfree-model-liveness.test.tsfails on the current pin once nano-30b leaves the picker. Merge this PR with a merge commit or a rebase, not a squash, so the pinned SHA stays on main; if it gets squashed, re-pin ClawRouter to the squash commit.🤖 Generated with Claude Code