Skip to content

fix(groq): replace the models api.groq.com retired - #593

Merged
di-sukharev merged 2 commits into
di-sukharev:masterfrom
tiltmeter-dev:fix/groq-retired-models
Nov 13, 2026
Merged

di-sukharev merged 2 commits into
di-sukharev:masterfrom
tiltmeter-dev:fix/groq-retired-models

Conversation

@torrresagus

Copy link
Copy Markdown
Contributor

What changed in the API

Notice groq-retired-models-2026-08 about api.groq.com's API describes this change, published 2026-06-17 and effective 2026-08-16. It is not signed by api.groq.com: it was read from a file the person running Tiltmeter trusts.

The notice's summary:

Groq shut down llama-3.3-70b-versatile and llama-3.1-8b-instant on August 16, 2026 for free and
developer-tier keys, after the Llama, Mixtral, Gemma, DeepSeek and Qwen models it retired between
2024 and 2026. Requests that name them return an error. Each change names Groq's recommended
replacement or, where that replacement was retired too, its current successor. Source:
https://console.groq.com/docs/deprecations

POST https://api.groq.com/openai/v1/chat/completions

Severity Change Path Before After
breaking value_retired $.model llama-3.1-8b-instant (no longer accepted), send openai/gpt-oss-20b instead
breaking value_retired $.model llama3-70b-8192 (no longer accepted), send openai/gpt-oss-120b instead
breaking value_retired $.model llama3-8b-8192 (no longer accepted), send openai/gpt-oss-20b instead
breaking value_retired $.model gemma2-9b-it (no longer accepted), send openai/gpt-oss-20b instead
breaking value_retired $.model gemma-7b-it (no longer accepted), send openai/gpt-oss-20b instead
breaking value_retired $.model llama-3.1-70b-versatile (no longer accepted), send openai/gpt-oss-120b instead
breaking value_retired $.model deepseek-r1-distill-llama-70b (no longer accepted), send openai/gpt-oss-120b instead

The notice's notes:

llama-3.1-8b-instant: Shut down on August 16, 2026: requests that name it return an error. The
shutdown applies to free and developer-tier keys; Groq enterprise customers with a committed-spend
contract keep the model. openai/gpt-oss-20b is Groq's recommended replacement. openai/gpt-oss-120b
and openai/gpt-oss-20b are reasoning models: they reason before they answer, and that reasoning
counts toward the completion token limit.
llama3-70b-8192: Shut down on August 30, 2025. Groq's replacement at the time,
llama-3.3-70b-versatile, was itself shut down on August 16, 2026; openai/gpt-oss-120b is its
recommended replacement.
llama3-8b-8192: Shut down on August 30, 2025. Groq's replacement at the time, llama-3.1-8b-instant,
was itself shut down on August 16, 2026; openai/gpt-oss-20b is its recommended replacement.
gemma2-9b-it: Shut down on October 8, 2025. Groq's replacement at the time, llama-3.1-8b-instant,
was itself shut down on August 16, 2026; openai/gpt-oss-20b is its recommended replacement.
gemma-7b-it: Shut down on December 18, 2024. Its replacement, gemma2-9b-it, was shut down on October
8, 2025, and that one's replacement on August 16, 2026; openai/gpt-oss-20b is the current successor.
llama-3.1-70b-versatile: Shut down on January 24, 2025. Its replacement, llama-3.3-70b-versatile,
was shut down on August 16, 2026; openai/gpt-oss-120b is the current successor.
deepseek-r1-distill-llama-70b: Shut down on October 2, 2025. openai/gpt-oss-120b is one of the two
replacements Groq named, and the one still served.

What this pull request changes

Groq has shut down the default model, llama3-70b-8192, and the other models in our Groq list, so requests naming them now fail. The Groq default and recommended model are now openai/gpt-oss-120b, and the Groq model list now offers Groq's named replacements, openai/gpt-oss-120b and openai/gpt-oss-20b. Duplicates are merged and llama-guard-3-8b is kept.

  • src/commands/config.ts
  • src/utils/errors.ts

Left for a human

  • Groq says openai/gpt-oss-120b and openai/gpt-oss-20b are reasoning models, and their reasoning counts toward the completion token limit. The configured OCO_TOKENS_MAX_OUTPUT may be too small for them, which could cut commit messages short. I didn't change it, because the right limit can't be known from the change list.
  • Users whose own config names a retired Groq model, such as OCO_MODEL=llama3-70b-8192, will still get errors until they change it. Only the defaults and the model list were updated.

This pull request was generated automatically by Tiltmeter from the unsigned notice groq-retired-models-2026-08 about api.groq.com's API. An AI agent (Claude Code with claude-opus-5-5) wrote these edits; review them like any other change.

Groq has shut down the default model, llama3-70b-8192, and the other models in our Groq list, so requests naming them now fail. The Groq default and recommended model are now openai/gpt-oss-120b, and the Groq model list now offers Groq's named replacements, openai/gpt-oss-120b and openai/gpt-oss-20b. Duplicates are merged and llama-guard-3-8b is kept.

Notice groq-retired-models-2026-08 about api.groq.com's API describes the change. It was read from a file and is not signed by the provider.
@di-sukharev

Copy link
Copy Markdown
Owner

please handle the new default as a reasoning model; it currently ignores the reasoning token limit.

Groq's openai/gpt-oss-* models, the new default and its replacement, are
reasoning models: their reasoning counts toward the completion limit. When
OCO_REASONING is not set, the Groq engine now sends max_completion_tokens
with OCO_REASONING_MAX_TOKENS instead of max_tokens and the sampling
parameters, and the diff splitting reserves that budget too. Other providers
send the same requests as before, and an explicit OCO_REASONING still wins.
@torrresagus

Copy link
Copy Markdown
Contributor Author

Thanks @di-sukharev, done in 62a8aeb. When OCO_REASONING is not set, the Groq engine now treats openai/gpt-oss-* (the new default and gpt-oss-20b) as reasoning models. Requests send max_completion_tokens with OCO_REASONING_MAX_TOKENS instead of max_tokens, temperature and top_p, and the diff splitting in generateCommitMessageByDiff reserves the same budget. Other providers send the same requests as before, and an explicit OCO_REASONING still wins.

I added a unit test for the Groq request. npm test (205 passing) and npm run typecheck pass locally.

@di-sukharev
di-sukharev merged commit 0441d61 into di-sukharev:master Nov 13, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants