Skip to content

v1.1.13

Choose a tag to compare

View all tags
@github-actions github-actions released this 04 Jul 15:47
· 35 commits to main since this release
v1.1.13
This tag was signed with the committer’s verified signature.
MitulShah1 Mitul Shah
GPG key ID: 9A023A4CFFD1D25D
Verified
Learn about vigilant mode.
47935ad
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

[1.1.13] — 2026-07-04

OpenAI-compatible provider fidelity release (part one). The fourth provider-readiness remediation phase aligns the Mistral, DeepSeek, Together, Fireworks, Cerebras, and Groq providers on request/response correctness, and adds two shared improvements that benefit every OpenAI-compatible provider. No breaking API changes relative to v1.1.12.

Fixed

  • Mistral sampling seed: the seed parameter is now sent as Mistral's random_seed field (Mistral ignores the standard seed), so a requested seed actually takes effect.
  • Mistral embedding dimensions: the embeddings dimensions parameter is now sent as Mistral's output_dimension field.
  • DeepSeek model list: deepseek-v4-flash and deepseek-v4-pro are now advertised, ahead of the retirement of deepseek-chat/deepseek-reasoner.
  • Groq model catalog: the two decommissioned models (mixtral-8x7b-32768, gemma2-9b-it) are removed and current production models (openai/gpt-oss-120b, openai/gpt-oss-20b) added.
  • DeepSeek error and finish_reason handling: chat errors now use the shared error envelope, and finish reasons are normalized to the canonical OpenAI vocabulary on both the streaming and non-streaming paths.

Changed

  • Streaming token usage (all OpenAI-compatible providers): streaming requests now set stream_options.include_usage, so providers that gate the terminal usage chunk on that flag report streaming token usage.
  • finish_reason normalization (all OpenAI-compatible providers): the shared chat and stream decoders normalize provider-specific finish reasons (e.g. Mistral's model_length → length) to the canonical OpenAI vocabulary.
  • Together default API domain (operator-visible): the default base URL is now https://api.together.ai (the current documented host) instead of the legacy https://api.together.xyz; deployments pinned to the old domain can still override it via TOGETHER_BASE_URL.
  • DeepSeek model discovery: DeepSeek now supports live /models discovery, so its advertised model list can self-update.
  • Shared base-URL validation: Mistral, DeepSeek, Together, Fireworks, Cerebras, and Groq now validate the configured base URL at construction, and Cerebras and Groq forward only the modern max_completion_tokens field.

Contributors

Thanks to everyone who shipped this release:

Full changelog: https://github.com/ferro-labs/ai-gateway/blob/v1.1.13/CHANGELOG.md

Contributors

MitulShah1
Assets 8
Loading