[Anandb71]

Summary

hey! nebius token factory removed eight serverless models on november 9, 2026, and two more nemotron ids leave on december 9, 2026. sim's nebius block still lists all of them, and the provider default (nebius/Qwen/Qwen3.5-397B-A17B) is one of the ones already gone, so a new nebius agent block starts on a model that no longer answers.

sources:

• november notice: https://docs.tokenfactory.nebius.com/november-2026-deprecation-notice
• december notice: https://docs.tokenfactory.nebius.com/model-deprecations-december-2026. the body says the models are removed "on December 9, 2026", while the note at the top says "On December 24, 2026: Affected Serverless models will no longer be available via API or Playground". i went with the earlier date
• nebius's public catalog, checked today: https://tokenfactory.nebius.com/api/public/models_info

what changed, all in the nebius block of apps/sim/providers/models.ts:

• defaultModel is now nebius/Qwen/Qwen3.8-27B, the replacement nebius recommends for Qwen3.5-397B-A17B. it's already in the list and has no sunset
• sunset: { status: 'deprecated' } on the eight ids that are no longer in the catalog: DeepSeek-V4.1-Flash, DeepSeek-V4-Flash-0731, DeepSeek-V4-Pro, Kimi-K2.6, Kimi-K2.7-Code, MiniCPM-V-4_5, Qwen3.5-397B-A17B and GLM-5.1. per the sunset doc comment, pickers hide them and pinned workflows still validate
• sunset: { status: 'legacy' } on NVIDIA-Nemotron-3-Nano-30B-A3B and nemotron-3-super-120b-a12b, which still answer until december 9 (nebius recommends nvidia/Nemotron-3_5-Lightning and nvidia/Nemotron-3.5-Super-VL), and on Qwen3-235B-A22B-Instruct-2507, which the catalog still serves but flags deprecated: true

once december 9 passes the two nemotrons can move to deprecated, happy to send that as a follow up.

i didn't add a nebius/nvidia/Nemotron-3.5-Super-VL entry. its price and context are public ($0.30 in, $0.90 out per 1M tokens, 1,048,576 context) but i couldn't confirm nativeStructuredOutputs without a key, and the add-model skill asks for every field to be verified, so i left it out instead of guessing. lightning is already in the list.

small heads up: apps/sim/providers/nebius/index.live.test.ts still uses nebius/openbmb/MiniCPM-V-4_5 for the vision case and nebius/Qwen/Qwen3.5-397B-A17B for the prompt fallback case, and both are gone now. i didn't touch the live test since i can't run it without a key.

no linked issue.

Product update (if relevant)

nebius users stop seeing retired models in the model picker, the two december nemotrons show as legacy with the upgrade prompt, and a new nebius agent block starts on a model that answers.

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation
• Other: ___________

Testing

• bun run --cwd apps/sim test providers/models.test.ts providers/nebius/index.test.ts providers/utils.test.ts: 244 passed. this includes the check that no provider default is sunset, which is why the default had to move
• bun run lint:check and bunx biome check apps/sim/providers/models.ts: clean
• bun run check:audits: 62 of 63 pass. the one failure is check:unused-exports (knip can't resolve @/.content/registry in apps/docs/app/[[...slug]]/page.tsx), and it fails the same way on untouched staging for me
• loaded the catalog with bun and read getModelSunsetStatus for every nebius id: 16 of the 24 stay in the picker, the new default has no sunset, and lookups work in lowercase too
• bun run type-check got killed for memory on my machine before it finished, so ci is the real check there. the new lines are the same sunset literal the file already uses for other providers
• no live nebius calls

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

no new tests: this is catalog data, and the existing models tests already cover the invariant that matters here (no provider default is sunset).

Screenshots/Videos

no ui code changed, only catalog data.

[vercel]

@Anandb71 is attempting to deploy a commit to the Sim Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] The changes affect Nebius model choices and the separate PII service's request handling.

The PR appears safe to merge, with a non-blocking concern about NER requests delaying unrelated PII endpoints.

Findings

1.  NER delays unrelated requests ▶

Summary

Updates Nebius model choices and reduces concurrent memory use in the PII service.

• Nebius blocks start on Qwen3.8-27B, and its older models get lifecycle labels.
• PII workers limit overlapping NER memory use.
• PII health checks answer while analysis waits for a worker.

Acknowledged deferrals from Anandb71: retired-model references in live tests remain unchanged because no API key was available; the December Nemotron retirement update is deferred until December 9; Nemotron-3.5-Super-VL is omitted because its structured-output support was not verified.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  Request[PII analysis request] --> Regex{Regex only?}
  Regex -->|Yes| Blank[Analyze without NER lock]
  Regex -->|No| Long{Over 100000 characters?}
  Long -->|Yes| Windows[Acquire NER lock per window]
  Long -->|No| Short[Acquire NER lock for text or batch]
  Windows --> Results[Return detected spans]
  Short --> Results
  Blank --> Results
  Health[Health request] --> Loop[Answer on event loop]

Loading

_{Reviews (1) · Last reviewed commit: "fix(providers): mark retired Nebius mode..." · Reviewed by Greptile}

[greptile-apps]

NER delays unrelated requests

Waiting for _NER_SLOT occupies a thread from the same pool used by regex-only analysis and anonymization. When concurrent NER requests fill that pool, those independent requests must also wait for NER work. Making /health async protects the probe but not these endpoints.

Consider moving the wait outside the shared threadpool or giving NER work a separate bounded pool.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!