Skip to content

2.23.0

Choose a tag to compare

View all tags
@Superagentic-AI Superagentic-AI released this 17 Oct 23:23
· 67 commits to main since this release
v2.23.0
14ec4bc

This release adds Gemini streaming, Gemma route fixes, vLLM support, Gemma on the Gemini API, and pricing and catalog updates.

Added

  • Stream CodeGem Gemini replies with streamGenerateContent (server-sent
    events), on by default and controlled by the stream setting. Text appears
    live while the stored model turn keeps every part and thought signature as
    received. Retries apply only before the first chunk, dropped streams resume
    from the last long decode checkpoint, and continuation hops stream too.
    Gemini 4 uses Google's 64K streaming output default.
  • Run CodeGem with Gemma 4 on the Gemini API (gemma-4-31b-it,
    gemma-4-26b-a4b-it, google provider), with Google's documented minimal
    and high thinking levels.
  • Run CodeGem local Gemma on vLLM, with the vLLM Gemma 4 recipe documented.
    With vLLM and llama.cpp, CodeGem reads the server's real context size.
  • Add gemini-3.1-flash-lite and gemini-3.1-pro-preview-customtools to the
    Google catalog.

Changed

  • Default change: Gemma 4 thinking is now off by default on local routes
    (Ollama, vLLM and llama.cpp). Turn it on with the reasoning setting.
  • Default change: CodeGem Gemini replies now stream by default. Set stream
    to false to restore single-response requests.
  • Local Gemma tool calls carry IDs and tool results are paired by
    tool_call_id and ordered to match the calls. Arguments are objects on
    Ollama and JSON strings on Chat Completions, as the OpenAI spec requires.
  • Gemma 4 tool-returned images stay inside the tool message instead of a
    synthetic user turn.
  • Gemma thinking is sent the same way on Ollama and Chat Completions routes
    and defaults to off, matching the Gemma 4 template and Ollama. Thinking from
    earlier user turns is no longer resent.
  • Gemma 3 system prompts are folded into the first user turn.
  • Gemma sampling uses Ollama's official defaults (temperature 1.0, top_p 0.95,
    top_k 64) only when none is set; Gemma 3 previously used temperature 0.2.

Fixed

  • Gemini cost estimates charge cached tokens at Google's published cache rates,
    apply the Gemini 3.1 Pro long-context tier, and follow scheduled price
    changes for Argon, Gemini 3.8 Flash and Gemini 3.6 Flash.
  • Correct the Gemini 3.6 Flash price to $0.75 input and $3.75 output per 1M tokens.

Security

Assets 4
Loading