Swift-native in-process engine
Model loading, generation, caching, and serving happen in one Swift process through Apple MLX. No Python service to install, supervise, or keep in sync.
v0.6.2available now
Your Mac.
Its models.
One native engine.
Run language and vision models through a native SwiftUI app, a real CLI, or a compatible API — all backed by the same Swift in-process MLX engine.
macOS 14+ · Apple Silicon · Apache 2.0
MacMLXCore
Swift-native
in-process
in-process
App
CLI
API
Inference stays on this Mac
~50 MBsingle app bundle
Defaultno Python runtime required
3surfaces, one core
100%local by default
01Why macMLX
Native is more than
a window frame.
The difference is underneath: the inference engine itself is Swift, runs inside the app process, and powers every interface.
One core, three first-class surfaces
Use the SwiftUI app, script with macmlx, or connect an existing OpenAI, Anthropic, Ollama, or MCP workflow. They share the same models and engine behavior.
Engine work you can inspect
macMLX ships prompt caching, model pooling, LoRA, VLM support, and a pure-Swift DeepSeek V3.2 port. Track G keeps tested results distinct from theoretical estimates so engine progress stays inspectable.
02What ships today
A small app with
a serious backend.
Everything below is available in v0.6.2 — not a concept render or a future promise.
01LOCAL
Native chat, from text to vision
Streaming Markdown, reasoning blocks, image attachments, conversation history, rewind, and per-model parameters in a focused macOS interface.
02API
Fits existing workflows
Compatible OpenAI, Anthropic, Ollama, MCP, embeddings, and rerank surfaces support agent and API tool loops plus structured output controls for supported workflows.
03ENGINE
Built for repeated work
Continuous batching, longest-common-prefix reuse, speculative decoding, and hot RAM / cold SSD cache tiers keep repeated and concurrent work practical.
04HUGGING FACE
Models without terminal archaeology
Browse, download, resume, update, and load MLX models in the app. Mirrors and commit-aware update detection are built in.
05CLI
A CLI that shares the core
Pull, run, serve, inspect, and stop from scripts or SSH without switching to a separate runtime.
06OBSERVE
Measure the machine you own
Benchmark throughput and first-token latency locally, then inspect XTC and KV-cache quantization behavior. Track G keeps tested measurements separate from theoretical estimates, with engine events available in Logs.
03Inside the engine
One Mac.
Five layers of control.
Follow one inference journey from Apple Silicon to sparse experts, shareable memory, adaptive admission, and controlled generation.
-
The memory is the architecture.
Unified memory keeps model weights, activations, and cache close to both CPU orchestration and integrated-GPU compute. MLX can move the work without building a second runtime around it.
-
Activate expertise, not every parameter.
A router selects a small expert set for each token, combines it with shared expertise, and merges the weighted result. macMLX ships a parity-verified pure-Swift DeepSeek V3.2 MoE path without claiming every MoE family.
-
Reuse the prompt. Branch only when needed.
Longest-common-prefix (LCP) reuse is available today across hot RAM and content-addressed SSD cold tiers. Paged allocation, block sharing, and copy-on-write remain planned for sharing prefixes across more concurrent work.
-
Admit work at the speed memory can sustain.
The fixed prefill admission is released and prevents a prompt wave from consuming every batch slot at once. Memory probes, pool limits, and eviction already shape admission; the unified adaptive guard remains future work.
-
Shape probability. Preserve reproducibility.
Temperature, top-p, XTC, and KV-cache quantization are available today. Top-k, min-p, presence, frequency, and repetition penalties, and per-request seed remain planned for a future control surface.
04Development progress
Shipping first.
Then going faster.
The website now separates released capability from active engineering work, so the roadmap reads as progress rather than promises.
Releasedv0.6.2
A capable local inference stack
- Language models + 14 detected VLM families
- Prompt cache, model pool, and LoRA adapters
- OpenAI, Anthropic, Ollama, MCP, embeddings, rerank
- Pure-Swift DeepSeek V3.2 architecture
Current releasev0.6.2 · Jul 11, 2026
The v0.6 engine is shipping
- Agent and API tool loops with structured output, XTC, and KV-cache quantization controls
- Continuous batching, LCP reuse, and speculative decoding runtime
- Track G distinguishes four tested models from theoretical InternLM3
- v0.6.1 hardening and v0.6.2 per-model chat-template overrides
05Get started
From download
to first token.
Install the app for the full macOS experience, or use the CLI when the terminal is already home.
$ macmlx pull mlx-community/Qwen3-8B-4bit $ macmlx run Qwen3-8B-4bit "Hello, Mac." $ macmlx serve
http://localhost:8000/v1ready