AFM — local AI infrastructure for Apple Silicon

Website · Documentation · GitHub releases

[!IMPORTANT] AFM is moving to a two-layer architecture built on AFMKit.

AFMKit is the reusable Swift foundation for building native AI apps and agents on Apple platforms. It owns the shared model contracts and provider runtimes for Apple Foundation Models, MLX, and DwarfStar, including generation, streaming, tool calling, structured output, and multimodal services.

maclocal-api remains the complete AFM product: the afm CLI, local OpenAI-compatible server, WebUI, packaging, and operational tooling. It now consumes one exact-versioned AFMKit dependency instead of carrying parallel provider implementations. This separation lets apps and agents embed AFMKit directly while server users continue to install and run AFM as before.

The migration is being delivered incrementally. Existing AFM commands and API compatibility remain the product contract while provider ownership moves into AFMKit and maclocal-api becomes a focused application layer on top of it.

What's new

AFM v0.9.18.1

AFM v0.9.18.1 is a focused stable hotfix over v0.9.18. It ensures stable release builds include the Apple Foundation Models provider and rejects release artifacts that do not link FoundationModels.framework.

v0.9.17 model highlights

AFM turns an Apple Silicon Mac into a private, OpenAI-compatible AI server. Run Hugging Face MLX models or Apple’s on-device Foundation Model, then connect the clients and SDKs you already use.

• Native Swift executable—no Python runtime for serving
• Local inference—no cloud account or API key
• Chat, streaming, tools, structured output, reasoning, and logprobs
• Vision OCR, speech, embeddings, and a built-in WebUI
• Prefix caching, concurrent decode, speculative decoding, and metrics
• Importable Swift packages for apps that need in-process inference

AFM is for Apple Silicon Macs running current macOS/Xcode toolchains. MLX model weights download from Hugging Face the first time you use them.

Install

[!NOTE] Stable v0.9.18.1 is the recommended release. It fixes stable packaging so the Apple Foundation Models provider is present and adds a release gate that rejects artifacts without FoundationModels.framework. Install afm-next only to preview changes made after v0.9.18.1.

v0.9.18.1 is a focused hotfix over v0.9.18. Use this stable release unless a newer nightly explicitly lists a post-v0.9.18.1 change you need.

Install a previous version

Older stable releases are kept as pinned formulae in the Homebrew tap and as version-pinned wheels on PyPI. This is useful for reproducing an issue against a specific build or rolling back without waiting for a new release.

Homebrew (pinned stable formulae): afm@<version> — available for 0.9.0, 0.9.1, and 0.9.3–0.9.10.

brew install scouzi1966/afm/afm@0.9.10
brew uninstall afm
brew link afm@0.9.10
afm --version

Homebrew (pinned nightly formulae): afm-next@<full-version> — for example, afm-next@0.9.15-next.20260808.e70cc52. See the Homebrew tap for available pinned nightlies.

brew install scouzi1966/afm/afm-next@0.9.15-next.20260808.e70cc52

pip (version-pinned wheels): install any published release by version.

pip install macafm==0.9.10
pip install --extra-index-url https://maclocal-ai.pages.dev/afm/wheels/simple/ \
  macafm-next==0.9.15.dev20260808

Start in two minutes

brew install scouzi1966/afm/afm

# Start a small MLX model and open the WebUI
afm mlx -m Qwen3-0.6B-4bit -w

AFM is now listening at http://127.0.0.1:9999/v1.

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 in one paragraph."}],
    "stream": false
  }'

Or use Apple’s on-device model:

afm -w

Native terminal chat

Use --tui when you want a private, full-screen chat without running an HTTP server:

# Apple Foundation Models
afm --tui

# Any supported MLX model (all normal sampling/runtime flags still apply)
afm mlx -m Qwen3-0.6B-4bit --tui

TUI changes have a model-free regression harness. make test-tui runs stable Markdown/math/code snapshots and exercises keyboard input, terminal sizing, alternate-screen cleanup, and raw-mode restoration through a real macOS pseudo-terminal. If an intentional visual change updates the expected output, run Scripts/test-tui.sh --record, inspect the snapshot diff, then rerun make test-tui normally. The same focused suite runs automatically on relevant pull requests.

The terminal UI streams responses, separates optional reasoning, and renders both answers and visible reasoning through a native CommonMark/GFM renderer. Headings, nested/task lists, quotes, tables, links, inline formatting, fenced code, and raw HTML are presented as inert terminal output. Code uses source-compiled Tree-sitter grammars for semantic highlighting and line numbers (details); unified diffs have distinct file, hunk, addition, and deletion styling. Inline and display LaTeX are rendered as readable Unicode math, including fractions, roots, super/subscripts, operators, Greek symbols, matrices, and cases. The UI also reports token and throughput statistics.

Reasoning is collapsed by default into a live activity row with its phase, animated cursor, elapsed time, and generated character count. Press Tab during generation to expand or collapse the reasoning panel without interrupting the model. Use /reasoning expanded, /reasoning collapsed, /reasoning off, or /reasoning last to control it explicitly.

It supports multiline editing, prompt history, cancellation, persisted/searchable sessions under ~/.afm/sessions, transcript export, attachments, terminal-width-aware tables, themes, and safe actions for response artifacts. Use /help for the command palette.

The navigation follows Codex CLI conventions. Normal chat output remains in Terminal scrollback. Press Ctrl+T to open the full transcript overlay, then scroll with a Mac trackpad or mouse wheel, arrows, Page Up/Down, Ctrl-U/Ctrl-D, or Home/End; press Ctrl+T again to close it. Add --no-alt-screen to keep overlays inline too. /blocks opens a session-wide code-block list: navigate with arrows or paging keys, press Enter for Copy/Save/Preview actions, and Escape to return. Numbered /save, /copy, and /open commands remain available for direct use.

Code is never executed automatically. /save refuses overwrites unless /save! is used, and only an explicit /open previews HTML or JavaScript in the browser. iTerm2 and Kitty can display local images inline; Terminal.app uses an explicit /image Quick Look fallback.

Choose your runtime

Model IDs without an organization default to mlx-community, so Qwen3-0.6B-4bit and mlx-community/Qwen3-0.6B-4bit both work. For stable, general-purpose MLX models, use mlx-community; experimental scouzi1966 MLX artifacts require an unreleased maclocal-api version and are described in Experimental scouzi1966 MLX models.

Evaluate a local model

AFM ships all 91 labeled variants from the repository's comprehensive MLX test as a deterministic, no-judge suite. The model loads once, every output and timing measurement is retained locally, and a self-contained HTML report opens when the run finishes.

afm mlx -m mlx-community/Qwen3-0.6B-4bit --eval

# Headless run, suite discovery, and custom-suite scaffolding
afm mlx -m <model> --eval --no-open
afm mlx --eval-list
afm mlx --eval-init my-suite
afm mlx --eval-validate ~/.afm/evals/my-suite.json
afm mlx -m <model> --eval-suite comprehensive --eval-suite my-suite

Run artifacts are stored in collision-safe ~/.afm/evals/<date-time>-<model>-<suite>/ directories. See Local model evaluations for the suite schema, deterministic checks, report contents, and security limits.

Why AFM works well for agents

AFM is built for multi-turn, tool-using clients—not only chat demos.

Pick a tool-calling mode

• Native (default): AFM detects the model’s own format and uses the narrowest parser. Use this for parity checks and benchmarks.
• Repair: add --tool-call-parser afm_adaptive_xml for JSON-in-XML fallback, type coercion, nullable-schema handling, and fuzzy tool-name matching. Add --fix-tool-args when a model renames arguments.
• Raw: add --tool-call-parser none to return the model’s tool markup as ordinary assistant content.

See MLX tool-calling modes for examples and benchmark guidance.

Connect an existing client

Most OpenAI-compatible clients need only a base URL and a placeholder API key:

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

Copy-ready guides:

OpenCode · OpenClaw · Cline · Continue · Aider · Cursor · Hermes

OpenClaw users can also generate a provider block directly:

afm mlx -m Qwen3-Coder-Next-4bit --openclaw-config

API surface

AFM also implements OpenAI-style file and batch-job endpoints under /v1/files and /v1/batches when the MLX batch service is active.

Apple-native tools

The CLI and HTTP server expose useful system frameworks without another service.

# OCR text or a table from an image/PDF
afm vision --file invoice.pdf --table

# Other Vision modes: text, table, barcode, classify, saliency, auto
afm vision --file photo.heic --mode classify --format json

# Speech recognition
afm speech transcribe --file meeting.wav --format srt

# Text to speech
afm speech synthesize "Hello from AFM" --voice nova --output hello.aac

# Dedicated OpenAI-compatible embeddings server (default port 9998)
afm embed

For vision-language models, add --vlm and pass one or more files with --media.

Performance controls

Defaults are a good starting point. Use these when the workload calls for them:

# Reuse prompt KV across requests
afm mlx -m <model> --enable-prefix-caching

# Save memory on long context
afm mlx -m <model> --kv-bits 8

# Fair-queue concurrent requests through one model
afm mlx -m <model> --concurrent 4

# Strict tool/JSON schemas with xgrammar
afm mlx -m <model> --enable-grammar-constraints

# Per-request device, memory, timing, and bandwidth estimates
afm mlx -m <model> --gpu-profile -s "Explain Metal kernels"

Supported checkpoints can also use speculative decoding:

• --mtp for compatible Qwen models. Qwen3.8 automatically prefetches the separately published MTP head matching the base checkpoint's quantization; use --mtp-model <repo-or-path> to override it.
• --eagle3 <drafter-directory> for supported dense Gemma4 models
• --dspark-support <support.gguf> for compatible DwarfStar DSpark workflows

Read decode optimizations before choosing a checkpoint or interpreting benchmark results.

Sampling and response controls

The MLX backend supports temperature, top_p, top_k, min_p, repetition_penalty, presence_penalty, seed, stop, logprobs, and top_logprobs.

Useful server defaults:

# Apply one JSON schema when requests omit response_format
afm mlx -m <model> \
  --guided-json '{"type":"object","properties":{"answer":{"type":"string"}},"required":["answer"]}' \
  --enable-grammar-constraints

# Disable model reasoning/thinking
afm mlx -m <model> --no-thinking

# Pin chat-template keyword arguments
afm mlx -m <model> --chat-template-kwargs '{"enable_thinking":false}'

Use AFMKit and AFM as Swift packages

The independent scouzi1966/AFMKit package defines the provider products:

• AFMKitCore — provider contracts and core types
• AFMOpenAICompat — OpenAI-compatible request/response types
• AFMKitMLX — MLX model loading and inference
• AFMKitApple — Apple Foundation Models backend
• AFMKitFoundationModelsMLX — macOS 27 MLX executor bridge
• AFMKitDwarfStar — DwarfStar runtime integration

This maclocal-api consumer package separately defines its application adapters:

• AFMKitFoundationModels — compatibility re-export of AFMKitApple
• AFMKitFoundationModels27 — macOS 27 application adapters
• AFMKitFoundationModels27DwarfStar — opt-in DwarfStar macOS 27 adapter
• AFMKitServices — vision, speech, and embedding services
• AFMKit — high-level headless inference facade
• AFMServer — Vapor HTTP layer
• afm — CLI executable

AFMKit is pinned to one exact public SwiftPM release so provider source, runtime resources, and the consumer contract advance together. See AFMKit URL consumption for the dependency and publication policy. Release tooling fails closed unless the manifest, lock, public tag, and anonymously fetchable revision all agree.

AFMKitFoundationModels27DwarfStar remains as a source-compatible re-export; the implementation and runtime lifetime management now live in AFMKit's AFMKitFoundationModelsDwarfStar product. This keeps AFM and independent AFMKit consumers on the same bridge while preserving a one-commit rollback.

Start with the AFMKit public API guide and the independent AFMKitCore consumer.

Build from source

git clone https://github.com/scouzi1966/maclocal-api.git
cd maclocal-api
./build.sh

The build resolves only the revisions in the tracked Package.resolved, initializes pinned submodules, builds the locked WebUI, and packages AFMKit-owned runtime resources without mutating dependency sources. AFMKit is public, so the normal build needs no repository token. Add --install to install under INSTALL_PREFIX (default /usr/local).

Requirements

• Apple Silicon Mac
• macOS 26 or newer for the complete feature set
• Xcode 27 for development builds
• Disk and unified memory appropriate for the model you choose

Small 0.6B–4B quantized models are the easiest way to confirm a setup. Large 30B-class models need substantially more unified memory.

Documentation map

• Client setup guides
• MLX tool calling
• Vision OCR API
• Embeddings API
• Images API
• Apple-native endpoints
• Model path resolution
• Decode optimizations
• vLLM observability and GuideLLM
• AFMKit public API
• Parameter combinations and use cases
• Supported model architecture catalog
• Roadmap

Contributing

Issues, reproducible test cases, documentation improvements, and model-compatibility reports are welcome. Read AGENTS.md and CLAUDE.md before changing build, test, or vendored integration code.

If AFM is useful to you, star the repository. You may also like Vesta AI Explorer, a full-featured native macOS AI app.

License

MIT