vmlx

---

JANG 2-bit destroys MLX 4-bit on MiniMax M2.5:

Quantization	MMLU (200q)	Size
JANG_2L (2-bit)	74%	89 GB
MLX 4-bit	26.5%	120 GB
MLX 3-bit	24.5%	93 GB
MLX 2-bit	25%	68 GB

Adaptive mixed-precision keeps critical layers at higher precision. Scores at jangq.ai. Models at JANGQ-AI.

---

Quickstart

Install from PyPI

Published on PyPI as vmlx -- install and run in one command:

# Recommended: uv (fast, no venv hassle)
brew install uv
uv tool install vmlx
vmlx serve mlx-community/Qwen3-8B-4bit

# Or: pipx (isolates from system Python)
brew install pipx
pipx install vmlx
vmlx serve mlx-community/Qwen3-8B-4bit

# Or: pip in a virtual environment
python3 -m venv ~/.vmlx-env && source ~/.vmlx-env/bin/activate
pip install vmlx
vmlx serve mlx-community/Qwen3-8B-4bit

Note: On macOS 14+, bare pip install fails with "externally-managed-environment". Use uv, pipx, or a venv.

The vMLX inference server is now running at http://0.0.0.0:8000 with an OpenAI + Anthropic compatible API. Works with any model from mlx-community -- thousands of models ready to go.

Or download the desktop app

Get MLX Studio -- a native macOS app with chat UI, model management, image generation, and developer tools. No terminal required. Just download the DMG and drag to Applications.

Use with OpenAI SDK

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
response = client.chat.completions.create(
    model="local",
    messages=[{"role": "user", "content": "Hello!"}],
    stream=True,
)
for chunk in response:
    print(chunk.choices[0].delta.content or "", end="", flush=True)

Use with Anthropic SDK

import anthropic

client = anthropic.Anthropic(base_url="http://localhost:8000/v1", api_key="not-needed")
message = client.messages.create(
    model="local",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
)
print(message.content[0].text)

Use with curl

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "local",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": true
  }'

---

Model Support

vMLX runs any MLX model. Point it at a HuggingFace repo or local path and go.

---

Features

Inference Engine

Distributed Inference (Multi-Mac)

Run models too large for a single Mac across 2+ machines. Each Mac loads a subset of transformer layers and they communicate hidden states over the network.

# On worker Macs:
pip install vmlx
vmlx-worker --secret mysecret

# On coordinator Mac (runs the server):
vmlx serve JANGQ-AI/Qwen3.5-Coder-Rerank-397B-A27B-JANG_2L --distributed --cluster-secret mysecret

5-Layer Cache Architecture

Request -> Tokens
    |
L1: Memory-Aware Prefix Cache (or Paged Cache)
    | miss
L2: Disk Cache (or Block Disk Store)
    | miss
Inference -> float16 KV states
    |
KV Quantization -> q4/q8 for storage
    |
Store back into L1 + L2

Tool Calling

Auto-detected parsers for every major model family:

qwen - llama - mistral - hermes - deepseek - glm47 - minimax - nemotron - granite - functionary - xlam - kimi - step3p5

Reasoning / Thinking Mode

Auto-detected reasoning parsers that extract <think> blocks:

qwen3 (Qwen3, QwQ, StepFun) - minimax_m2 (MiniMax M2/M2.5/M2.7) - deepseek_r1 (DeepSeek R1, Gemma 3, GLM, Phi-4) - openai_gptoss (GLM Flash, GPT-OSS)

Audio

---

Image Generation & Editing

Generate and edit images locally with Flux models via mflux.

pip install vmlx[image]

# Image generation
vmlx serve schnell                    # or dev, z-image-turbo
vmlx serve ~/.mlxstudio/models/image/flux1-schnell-4bit

# Image editing
vmlx serve qwen-image-edit            # instruction-based editing

Generation API

curl http://localhost:8000/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
    "model": "schnell",
    "prompt": "A cat astronaut floating in space with Earth in the background",
    "size": "1024x1024",
    "n": 1
  }'

# Python (OpenAI SDK)
response = client.images.generate(
    model="schnell",
    prompt="A cat astronaut floating in space",
    size="1024x1024",
    n=1,
)

Editing API

# Edit an image with a text prompt (Flux Kontext / Qwen Image Edit)
curl http://localhost:8000/v1/images/edits \
  -H "Content-Type: application/json" \
  -d '{
    "model": "flux-kontext",
    "prompt": "Change the background to a sunset",
    "image": "<base64-encoded-image>",
    "size": "1024x1024",
    "strength": 0.8
  }'

# Python
import base64
with open("source.png", "rb") as f:
    image_b64 = base64.b64encode(f.read()).decode()

response = requests.post("http://localhost:8000/v1/images/edits", json={
    "model": "flux-kontext",
    "prompt": "Make the sky purple",
    "image": image_b64,
    "size": "1024x1024",
    "strength": 0.8,
})

Supported Image Models

Generation Models:

Editing Models:

---

API Reference

API Gateway

The desktop app runs an API Gateway on a single port (default 8080) that routes requests to all loaded models by name. Run multiple models simultaneously and access them all through one URL.

# All models accessible through the gateway
curl http://localhost:8080/v1/chat/completions \
  -d '{"model": "Qwen3.5-122B", "messages": [{"role": "user", "content": "Hi"}]}'

# Works with Ollama CLI too
OLLAMA_HOST=http://localhost:8080 ollama run Qwen3.5-122B

The gateway supports OpenAI, Anthropic, and Ollama wire formats. Configure the port in the API tab.

Endpoints

OpenAI / Anthropic

Ollama

curl Examples

Chat completion (streaming)

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "local",
    "messages": [{"role": "user", "content": "Explain quantum computing in 3 sentences."}],
    "stream": true,
    "temperature": 0.7
  }'

Chat completion with thinking mode

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "local",
    "messages": [{"role": "user", "content": "Solve: what is 23 * 47?"}],
    "enable_thinking": true,
    "stream": true
  }'

Tool calling

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "local",
    "messages": [{"role": "user", "content": "What is the weather in Tokyo?"}],
    "tools": [{
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather for a location",
        "parameters": {
          "type": "object",
          "properties": {
            "location": {"type": "string", "description": "City name"}
          },
          "required": ["location"]
        }
      }
    }]
  }'

Anthropic Messages API

curl http://localhost:8000/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: not-needed" \
  -H "anthropic-version: 2023-06-01" \
  -d '{
    "model": "local",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Embeddings

curl http://localhost:8000/v1/embeddings \
  -H "Content-Type: application/json" \
  -d '{
    "model": "local",
    "input": "The quick brown fox jumps over the lazy dog"
  }'

Text-to-speech

curl http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "kokoro",
    "input": "Hello, welcome to vMLX!",
    "voice": "af_heart"
  }' --output speech.wav

Speech-to-text

curl http://localhost:8000/v1/audio/transcriptions \
  -F file=@audio.wav \
  -F model=whisper

Image generation

curl http://localhost:8000/v1/images/generations \
  -H "Content-Type: application/json" \
  -d '{
    "model": "schnell",
    "prompt": "A mountain landscape at sunset",
    "size": "1024x1024"
  }'

Reranking

curl http://localhost:8000/v1/rerank \
  -H "Content-Type: application/json" \
  -d '{
    "model": "local",
    "query": "What is machine learning?",
    "documents": [
      "ML is a subset of AI",
      "The weather is sunny today",
      "Neural networks learn from data"
    ]
  }'

Cache stats

curl http://localhost:8000/v1/cache/stats

Health check

curl http://localhost:8000/health

---

Desktop App

vMLX includes a native macOS desktop app (MLX Studio) with 5 modes:

Download

Get the latest DMG from MLX Studio Releases, or build from source:

git clone https://github.com/jjang-ai/vmlx.git
cd vmlx/panel
npm install && npm run build
npx electron-builder --mac dmg

Menu Bar

vMLX lives in your menu bar showing all running models, GPU memory usage, and quick controls.

---

Advanced Quantization

vMLX supports standard MLX quantization (4-bit, 8-bit uniform) out of the box. For users who want to push further, JANG adaptive mixed-precision assigns different bit widths to different layer types -- attention gets more bits, MLP layers get fewer -- achieving better quality at the same model size.

JANG Profiles

Convert

pip install vmlx[jang]

# Standard MLX quantization
vmlx convert my-model --bits 4

# JANG adaptive quantization
vmlx convert my-model --jang-profile JANG_3M

# Activation-aware calibration (better at 2-3 bit)
vmlx convert my-model --jang-profile JANG_2L --calibration-method activations

# Serve the converted model
vmlx serve ./my-model-JANG_3M --continuous-batching --use-paged-cache

Pre-quantized JANG models are available at JANGQ-AI on HuggingFace.

Smelt Mode (Partial Expert Loading)

For MoE models that don't fit in RAM, Smelt loads only a subset of experts per layer from SSD and keeps the backbone resident. Routing is biased toward the resident experts, so response quality stays coherent while RAM usage drops. Trade-off: throughput scales inversely with expert % loaded, because expert swaps hit SSD on the hot path.

# Load 50% of experts per layer (default when --smelt alone)
vmlx serve ./MyMoE-JANG_4M --smelt --smelt-experts 50

# Aggressive: load 25% — smallest RAM, slowest
vmlx serve ./MyMoE-JANG_4M --smelt --smelt-experts 25

Benchmarks on Nemotron-Cascade-2-30B-A3B-JANG_4M (23 MoE layers × 128 experts, Apple M3 Ultra / 128 GB, dedicated machine, no parallel models):

* Responses too short (2-5 tokens) for reliable steady-state tok/s measurement at 25 %. Subjectively responsive.

All three configurations produced coherent, non-looping output. No quality degradation observed — routing bias keeps the model on-topic.

Credit: Smelt mode is inspired by Anemll's flash-moe — a pure C / Objective‑C / Metal inference engine that showed huge MoE models (Qwen3.5-397B) can run on modest Apple Silicon hardware by streaming expert weights from SSD with pread() on demand. vMLX Smelt takes a different implementation path: Python/MLX, tied to the JANG quantization format, and loading a fixed subset of experts per layer at startup (backbone resident, routing biased toward the loaded subset) rather than on-demand per-token. It plugs into the full vMLX server with continuous batching, paged cache, and OpenAI-compatible API. Different engine, same core insight — thanks to the flash-moe team for validating the approach.

Smelt is mutually exclusive with VLM mode. vMLX detects smelt and automatically disables --is-mllm (with a warning) because the vision tower is not wired through the partial-expert loader — image input on a smelt-loaded VLM would produce garbage logits. Use a text-only model when running smelt, or disable smelt when running a VLM.

Smelt requires an MoE model in JANG format. Not compatible with dense models (no experts to partial-load) or with non-JANG formats.

---

CLI Commands

vmlx serve <model>              # Start inference server
vmlx convert <model> --bits 4   # MLX uniform quantization
vmlx convert <model> -j JANG_3M # JANG adaptive quantization
vmlx info <model>               # Model metadata and config
vmlx doctor <model>             # Run diagnostics
vmlx bench <model>              # Performance benchmarks
vmlx-worker --secret <secret>   # Start distributed worker node

---

Configuration

Server Options

vmlx serve <model> \
  --host 0.0.0.0 \              # Bind address (default: 0.0.0.0)
  --port 8000 \                 # Port (default: 8000)
  --api-key sk-your-key \       # Optional API key authentication
  --continuous-batching \       # Enable concurrent request handling
  --enable-prefix-cache \       # Reuse KV states for repeated prompts
  --use-paged-cache \           # Block-based KV cache with dedup
  --kv-cache-quantization q8 \  # Quantize cache: q4 or q8
  --enable-disk-cache \         # Persist cache to SSD
  --enable-jit \                # JIT Metal kernel compilation
  --tool-call-parser auto \     # Auto-detect tool call format
  --reasoning-parser auto \     # Auto-detect thinking format
  --log-level INFO \            # Logging: DEBUG, INFO, WARNING, ERROR
  --max-model-len 8192 \        # Max context length
  --speculative-model <model> \ # Draft model for speculative decoding
  --enable-pld \                # Prompt Lookup Decoding — no draft model, best for code/JSON/schemas
  --distributed \               # Enable multi-Mac pipeline parallelism
  --cluster-secret <secret> \   # Shared auth secret for workers
  --distributed-mode pipeline \ # pipeline (default) or tensor (coming soon)
  --worker-nodes ip:port,... \  # Manual worker IPs (overrides auto-discovery)
  --cors-origins "*"            # CORS allowed origins

Quantization Options

vmlx convert <model> \
  --bits 4 \                    # Uniform quantization bits: 2, 3, 4, 6, 8
  --group-size 64 \             # Quantization group size (default: 64)
  --output ./output-dir \       # Output directory
  --jang-profile JANG_3M \      # JANG mixed-precision profile
  --calibration-method activations  # Activation-aware calibration

Image Generation & Editing Options

pip install vmlx[image]

# Generation models
vmlx serve schnell \            # or dev, z-image-turbo
  --image-quantize 4 \          # Quantization: 4, 8 (omit for full precision)
  --port 8001

# Editing models
vmlx serve qwen-image-edit \    # Instruction-based editing (full precision only)
  --port 8001

# Local model directory
vmlx serve ~/.mlxstudio/models/image/FLUX.1-schnell-mflux-4bit

Audio Options

TTS and STT require the mlx-audio package:

pip install mlx-audio

# TTS: serve Kokoro model
vmlx serve kokoro --port 8002

# STT: serve Whisper model
vmlx serve whisper --port 8003

Optional Dependencies

pip install vmlx              # Core: text LLMs, VLMs, embeddings, reranking
pip install vmlx[image]       # + Image generation (mflux)
pip install vmlx[jang]        # + JANG quantization tools
pip install vmlx[dev]         # + Development/testing tools
pip install vmlx[image,jang]  # Multiple extras

---

Architecture

+--------------------------------------------+
|          Desktop App (Electron)             |
|   Chat | Server | Image | Tools | API      |
+--------------------------------------------+
|          Session Manager (TypeScript)       |
|   Process spawn | Health monitor | Tray     |
+--------------------------------------------+
|         vMLX Engine (Python / FastAPI)       |
|  +--------+  +---------+  +-----------+    |
|  |Simple  |  | Batched |  | ImageGen  |    |
|  |Engine  |  | Engine  |  | Engine    |    |
|  +---+----+  +----+----+  +-----+-----+    |
|      |            |              |          |
|  +---+------------+--+    +-----+-----+    |
|  | mlx-lm / mlx-vlm  |    |  mflux    |    |
|  +--------+-----------+    +-----------+    |
|           |                                 |
|  +--------+----------------------------+    |
|  |       MLX Metal GPU Backend          |    |
|  | quantized_matmul | KV cache | SDPA   |    |
|  +--------------------------------------+    |
+--------------------------------------------+
|  L1: Prefix Cache (Memory-Aware / Paged)    |
|  L2: Disk Cache (Persistent / Block Store)  |
|  KV Quant: q4/q8 at storage boundary       |
+--------------------------------------------+

---

Contributing

Contributions are welcome. Here is how to set up a development environment:

git clone https://github.com/jjang-ai/vmlx.git
cd vmlx

# Python engine
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev,jang,image]"
pytest tests/ -k "not Async"    # 2000+ tests

# Electron desktop app
cd panel && npm install
npm run dev                      # Development mode with hot reload
npx vitest run                   # 1545+ tests

Project Structure

vmlx/
  vmlx_engine/          # Python inference engine (FastAPI server)
  panel/                # Electron desktop app (React + TypeScript)
    src/main/           # Electron main process
    src/renderer/       # React frontend
    src/preload/        # IPC bridge
  tests/                # Python test suite
  assets/               # Screenshots and logos

Guidelines

• Run the full test suite before submitting PRs
• Follow existing code style and patterns
• Include tests for new features
• Update documentation for user-facing changes

---

License

Apache License 2.0 -- see LICENSE.

---

---

한국어 (Korean)

vMLX — Apple Silicon을 위한 로컬 AI 엔진

Mac에서 LLM, VLM, 이미지 생성 및 편집 모델을 완전히 로컬로 실행하세요. OpenAI + Anthropic 호환 API. 클라우드 없음. API 키 불필요. 데이터가 기기를 떠나지 않습니다.

빠른 시작

pip install vmlx
vmlx serve mlx-community/Llama-3.2-3B-Instruct-4bit

주요 기능

이미지 생성

pip install vmlx[image]
vmlx serve schnell          # 빠른 생성 (4 단계)
vmlx serve dev              # 고품질 생성 (20 단계)

이미지 편집

vmlx serve qwen-image-edit  # 텍스트 지시 기반 이미지 편집

# 이미지 편집 API
curl http://localhost:8000/v1/images/edits \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen-image-edit",
    "prompt": "배경을 해질녘으로 변경",
    "image": "<base64 인코딩된 이미지>",
    "size": "1024x1024",
    "strength": 0.8
  }'

데스크톱 앱 (MLX Studio)

macOS 네이티브 데스크톱 앱으로 5가지 모드를 제공합니다:

설치

pip install vmlx              # 기본: 텍스트 LLM, VLM, 임베딩
pip install vmlx[image]       # + 이미지 생성/편집 (mflux)
pip install vmlx[jang]        # + JANG 양자화 도구
pip install vmlx[audio]       # + TTS/STT (mlx-audio)

라이선스

Apache License 2.0 — LICENSE 참조.

---