Skip to content

afm 0.9.14

Latest

Choose a tag to compare

@scouzi1966 scouzi1966 released this 17 Jul 12:11

afm 0.9.14

Apple Foundation Models + MLX local models — OpenAI-compatible API, WebUI, all Swift.

Changes since v0.9.13

  • Fix concurrent cache and grammar decoding (419ac81)
  • Fix pairwise assertions for thinking models (0cd5d5c)
  • Fix duplicate streaming tool calls (e98da7b)
  • Restore tool parser auto-detection (eb0e623)
  • Update stable release to 0.9.14 (549c576)
  • Fix merged MLX regressions (60d4cc8)
  • Fix Cohere raw output tag preservation (1ab78bf)
  • Merge remote-tracking branch 'origin/main' into codex/pr151-raw-mode (0d99a0f)
  • Merge pull request #150 from scouzi1966/fix/gemma4-vision-processor (bcb940f)
  • Merge pull request #147 from scouzi1966/fix/gemma4-eseries-quantized-scaledlinear (a6f6c62)
  • Merge remote-tracking branch 'origin/main' into fix/qwen36-malformed-toolcall-parser (ec16742)
  • fix(build): drop swift-jinja <2.4.0 cap (2.4.1 restored transformers compat) (885daf3)
  • Extract Cohere (cohere2_moe) reasoning into reasoning_content (3e10ca2)
  • fix(vlm): map gemma4 to Gemma4Processor in processorTypeOverrides (#149) (1f7a9be)
  • fix(mlx): support quantized per_layer_model_projection in Gemma 4 E-series (3769451)
  • fix(mlx): address round-2 review — parallel calls, brace fidelity, repair gating, JSON nulls (5d265d7)
  • fix(mlx): raise implicit max_tokens fallback 512 -> 4096 (2ff91b7)
  • fix(mlx): honor raw mode end-to-end + quote-aware malformed field parser (review P1s) (8b2ed64)
  • fix(build): make branch compile from a clean clone (review P0) (e56bd18)
  • Add nightly test results for 2026-07-02 (33e60dd) (7aeaf6a)
  • Update README nightly references to 20260703-33e60dd (413b7f9)
  • fix(build): plist check false negative under pipefail (grep -q SIGPIPE) (33e60dd)
  • Merge pull request #143 from scouzi1966/feature/qmv-wide-backport (b1e7b5c)
  • fix(mlx): gate qmv_wide dispatch to validated bit widths (4/8) (8d83512)
  • perf(mlx): backport qmv_wide small-batch quantized matvec (mlx#3764) (c24e828)
  • fix(cli): run async utility commands off the main actor (ce5fde9)
  • Vendor VulcanBench test fixture (7cbda3a)
  • Fix Qwen MLX tool-call parity path (059ce11)
  • Cancel abandoned non-streaming MLX generations (337e3a1)
  • Align MLX tool prompts with Python server (af75e41)
  • Repair collapsed Qwen tool arguments (231c680)
  • Fix Qwen tool-call parsing without parser override (6458cd7)
  • test(agentblaster): add reusable AgentBlaster × afm benchmark harness (7d70a7e)
  • docs(skill): note maintainer now has a TestPyPI account for wheel validation (4555b77)
  • chore: bump development version to 0.9.14 after v0.9.13 release (3f4fcf9)
  • docs(README): note that stable v0.9.13 is the absolute latest version (70f7c76)
  • docs(README): bump stable references to v0.9.13 (e779787)

Install / Upgrade via Homebrew

Fresh install:

brew tap scouzi1966/afm
brew install scouzi1966/afm/afm

Upgrade:

brew upgrade afm

Install via PyPI

pip install macafm==0.9.14