Vesta 0.9.7
MLX backend
- New: Gemma 4 family —
mlx-community/gemma-4-26b-a4b-it-4bit,gemma-4-e2b-it-4bit,gemma-4-e4b-it-4bit,gemma-4-31b-it-4bit, etc. Model typesgemma4andgemma4_textare now recognized; verified end-to-end ongemma-4-26b-a4b-it-4bit(chat template renders, generation @ 49 tok/s on M-series). - New: Qwen 3.6 family —
qwen3_6,qwen3_6_moe. Plus aliasing for the dotted variants (qwen3.5,qwen3.5_moe,qwen3.5_next,qwen3.5_vl,qwen3.6,qwen3.6_moe,qwen3_6_next,qwen3_6_vl,qwen3_5_vl). - Updated maclocal-api submodule — picks up MLX-Swift patches:
Gemma4Text.swift,Gemma4VLM.swift,Gemma4FunctionParser.swift, plus VLM/LLM factory updates. - swift-jinja 2.1.1 → 2.3.5 — fixes a parser bug where literal
{characters intermixed with{%- %}blocks (used by Gemma 4's chat template for JSON-like tool/output formatting) tripped the lexer. Without this, Gemma 4 chat fails withJinja.JinjaError error 1.
Build / packaging
- Fix
Scripts/mlx-audio-vesta.patch: the swift-transformers version line was authored as a diff context line instead of a-/+change;--from-scratchbuilds now resolve the SPM graph cleanly (mlx-audio + mlx-swift-lm both atfrom: "1.3.0"). - CFBundleVersion bumped to 10.
Install
brew tap scouzi1966/afm
brew install --cask scouzi1966/afm/vesta-macDownload
Vesta-0.9.7.dmg
SHA256: d6b488f24ddbeddfeda016c165ffa8dadd99a853774561663e0291a5051f7bbe