AFM 0.9.19
What's new since 0.9.18
- Expanded model support and performance: Qwen 3.8 Flash Next, GLM 5.3 Flash, and DeepSeek V4 Flash 0731 improvements, including MTP decoding, Qwen Next execution optimizations, GLM short-block sparse verification, and bounded-memory fixes for GLM and DeepSeek.
- Native checkpoint conversion: AFM conversion workflows for supported models, with Qwen Next conversion controls, mapped or resident n-gram storage, and improved storage preflight and resume handling.
- More API compatibility: OpenAI Responses, Anthropic Messages and token counting, and OpenAI-compatible image endpoints for supported backends. Improved tool-call collection, assistant prefill handling, request validation, and rate limiting.
- Cache and runtime reliability: Improvements to prefix-cache validation, batch dispatch, model capability reporting, and fatal startup error handling.
- Updated WebUI and packaging: Updated bundled WebUI, disabled the runtime-control side panel pending redesign, and strengthened Homebrew/Python resource checks and Foundation Models build safeguards.
- Better evaluation and observability: Feature-specific test prompts, full prompt and per-test AI-judge reporting, separate native-conformance/model-behavior/forced-parser results, and improved timing and failure attribution.
Install
brew update
brew install scouzi1966/afm/afm
# Existing stable installation:
brew upgrade scouzi1966/afm/afm
# Switching from staging: brew unlink afm-staging before installing stable.afm --version reports v0.9.19, with no nightly suffix. An Apple Silicon/macOS 26+ Python wheel is also included in Assets.
Test reports
Download afm-v0.9.19-test-reports.tar.gz from Assets for detailed qualification results, known limitations, HTML reports, annotated Context charts, supporting raw data, and checksum inventories. The archive includes the September 8 release-candidate results and Qwen 3.8 27B installed-binary tests on M3 Ultra and M4 Pro.