AFM — local AI infrastructure for Apple Silicon

AFM is a native executable for Apple Silicon. The shortest path is Homebrew and a small MLX checkpoint that verifies the complete request path.

Requirements

Apple Silicon Mac.
macOS 26 or newer for the complete feature set.
Xcode 27 for development builds.
Enough disk and unified memory for the checkpoint you choose.

1. Install

brew install scouzi1966/afm/afm

# Nightly builds
brew install scouzi1966/afm/afm-next

# Python packaging is also available; serving remains native
pip install macafm

To build from source, clone the repository and run ./build.sh. The script initializes submodules, applies AFM-owned vendor patches, builds WebUI assets and Metal resources, then creates the release executable.

2. Choose model storage

# Standard Hugging Face download cache
export HF_HUB_CACHE=~/.cache/huggingface/hub

# Optional read-only lookup root for your curated side-loaded models
export MACAFM_MLX_MODEL_CACHE=/Volumes/Models/afm

3. Choose a runtime

Runtime	Use it for	Start command	Boundary
MLX	Open models, VLMs, agent controls, tuning	`afm mlx -m <model>`	Weights download from Hugging Face or resolve locally.
Foundation Models	Apple's system model and `.fmadapter` LoRA	`afm`	System availability and supported OS determine capability.
DwarfStar	Compatible fixed-schedule Metal checkpoints	`--mlx-runtime dwarfstar`	Only checkpoints prepared for its layout contract.
Gateway	One model list for other local servers	`afm --gateway`	Execution stays in Ollama, LM Studio, Jan, or the discovered backend.

Unqualified model IDs default to the mlx-community organization. For example, Qwen3-0.6B-4bit and mlx-community/Qwen3-0.6B-4bit resolve to the same repository.

4. Start a small model

afm mlx -m Qwen3-0.6B-4bit -w

-w opens the built-in WebUI. The API listens at http://127.0.0.1:9999/v1. See the WebUI guide for attachments, model routing, metrics, and limitations.

5. Make a request

curl http://127.0.0.1:9999/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "Qwen3-0.6B-4bit",
    "messages": [{"role":"user","content":"Explain unified memory briefly."}],
    "stream": false
  }'

6. Connect a client

Most OpenAI-compatible clients need only these values:

Base URL: http://127.0.0.1:9999/v1
API key:  x

AFM does not gate localhost access on an API key; use a placeholder where a client requires one. Copy-ready source guides exist for OpenCode, OpenClaw, Cline, Continue, Aider, Cursor, and Hermes.

# Generate an OpenClaw provider block with detected capabilities
afm mlx -m Qwen3-Coder-Next-4bit --openclaw-config

What to configure next

For long, stable agent prompts: add --enable-prefix-caching.
For multiple client sessions: add --concurrent N.
For strict token-level JSON: add --enable-grammar-constraints.
For a VLM: add --vlm and pass media with --media.