LLM inference, optimized for your Mac
Continuous batching and tiered KV caching, managed directly from your menu bar.
This fork has support for the Gemma 4 models with tool calls and channel thinking parsing. Validated using Gemma-4-31B-it with Pi as the testing harnesses.
Validation Testing:
Running pi --tools read -ne -ns --model gemma-4-31b-it -m "Call read for README.md"
Current Limitations:
Enabling thinking with tools results in some form of generation coruption. Been unable to determine why this is ocurring with model generating rambling "// a single tool call" and the like output. During some runs it does generate some sensical output but something is wrong with how message template is applied.
Install · Quickstart · Features · Models · CLI Configuration · Benchmarks · oMLX.ai
Every LLM server I tried made me choose between convenience and control. I wanted to pin everyday models in memory, auto-swap heavier ones on demand, set context limits - and manage it all from a menu bar.
oMLX persists KV cache across a hot in-memory tier and cold SSD tier - even when context changes mid-conversation, all past context stays cached and reusable across requests, making local LLMs practical for real coding work with tools like Claude Code. That's why I built it.
git clone https://github.com/jundot/omlx.git
cd omlx
pip install -e . # Core only
pip install -e ".[mcp]" # With MCP (Model Context Protocol) support
python -m omlx.cli serveRequires macOS 15.0+ (Sequoia), Python 3.10+, and Apple Silicon (M1/M2/M3/M4).
Supports text LLMs, vision-language models (VLM), OCR models, embeddings, and rerankers on Apple Silicon.
Web UI at /admin for real-time monitoring, model management, chat, benchmark, and per-model settings. Supports English, Korean, Japanese, and Chinese. All CDN dependencies are vendored for fully offline operation.
Run VLMs with the same continuous batching and tiered KV cache stack as text LLMs. Supports multi-image chat, base64/URL/file image inputs, and tool calling with vision context. OCR models (DeepSeek-OCR, DOTS-OCR, GLM-OCR) are auto-detected with optimized prompts.
Block-based KV cache management inspired by vLLM, with prefix sharing and Copy-on-Write. The cache operates across two tiers:
- Hot tier (RAM): Frequently accessed blocks stay in memory for fast access.
- Cold tier (SSD): When the hot cache fills up, blocks are offloaded to SSD in safetensors format. On the next request with a matching prefix, they're restored from disk instead of recomputed from scratch - even after a server restart.
Handles concurrent requests through mlx-lm's BatchGenerator. Prefill and completion batch sizes are configurable.
Context scaling support for running smaller context models with Claude Code. Scales reported token counts so that auto-compact triggers at the right timing, and SSE keep-alive prevents read timeouts during long prefill.
Load LLMs, VLMs, embedding models, and rerankers within the same server. Models are managed through a combination of automatic and manual controls:
- LRU eviction: Least-recently-used models are evicted automatically when memory runs low.
- Manual load/unload: Interactive status badges in the admin panel let you load or unload models on demand.
- Model pinning: Pin frequently used models to keep them always loaded.
- Per-model TTL: Set an idle timeout per model to auto-unload after a period of inactivity.
- Process memory enforcement: Total memory limit (default: system RAM - 8GB) prevents system-wide OOM.
Configure sampling parameters, chat template kwargs, TTL, model alias, model type override, and more per model directly from the admin panel. Changes apply immediately without server restart.
- Model alias: set a custom API-visible name.
/v1/modelsreturns the alias, and requests accept both the alias and directory name. - Model type override: manually set a model as LLM or VLM regardless of auto-detection.
Chat directly with any loaded model from the admin panel. Supports conversation history, model switching, dark mode, reasoning model output, and image upload for VLM/OCR models.
Search and download MLX models from HuggingFace directly in the admin dashboard. Browse model cards, check file sizes, and download with one click.
Set up OpenClaw, OpenCode, and Codex directly from the admin dashboard with a single click. No manual config editing required.
One-click benchmarking from the admin panel. Measures prefill (PP) and text generation (TG) tokens per second, with partial prefix cache hit testing for realistic performance numbers.
Native PyObjC menubar app (not Electron). Start, stop, and monitor the server without opening a terminal. Includes persistent serving stats (survives restarts), auto-restart on crash, and in-app auto-update.
Drop-in replacement for OpenAI and Anthropic APIs. Supports streaming usage stats (stream_options.include_usage), Anthropic adaptive thinking, and vision inputs (base64, URL).
| Endpoint | Description |
|---|---|
POST /v1/chat/completions |
Chat completions (streaming) |
POST /v1/completions |
Text completions (streaming) |
POST /v1/messages |
Anthropic Messages API |
POST /v1/embeddings |
Text embeddings |
POST /v1/rerank |
Document reranking |
GET /v1/models |
List available models |
Supports all function calling formats available in mlx-lm, JSON schema validation, and MCP tool integration. Tool calling requires the model's chat template to support the tools parameter. The following model families are auto-detected via mlx-lm's built-in tool parsers:
| Model Family | Format |
|---|---|
| Llama, Qwen, DeepSeek, etc. | JSON <tool_call> |
| Qwen3.5 Series | XML <function=...> |
| Gemma | <start_function_call> |
| GLM (4.7, 5) | <arg_key>/<arg_value> XML |
| MiniMax | Namespaced <minimax:tool_call> |
| Mistral | [TOOL_CALLS] |
| Kimi K2 | <|tool_calls_section_begin|> |
| Longcat | <longcat_tool_call> |
Models not listed above may still work if their chat template accepts tools and their output uses a recognized <tool_call> XML format. For tool-enabled streaming, assistant text is emitted incrementally while known tool-call control markup is suppressed from visible content; structured tool calls are emitted after parsing the completed turn.
Point --model-dir at a directory containing MLX-format model subdirectories. Two-level organization folders (e.g., mlx-community/model-name/) are also supported.
~/models/
├── Step-3.5-Flash-8bit/
├── Qwen3-Coder-Next-8bit/
├── gpt-oss-120b-MXFP4-Q8/
├── Qwen3.5-122B-A10B-4bit/
└── bge-m3/
Models are auto-detected by type. You can also download models directly from the admin dashboard.
| Type | Models |
|---|---|
| LLM | Any model supported by mlx-lm |
| VLM | Qwen3.5 Series, GLM-4V, Pixtral, and other mlx-vlm models |
| OCR | DeepSeek-OCR, DOTS-OCR, GLM-OCR |
| Embedding | BERT, BGE-M3, ModernBERT |
| Reranker | ModernBERT, XLM-RoBERTa |
# Memory limit for loaded models
omlx serve --model-dir ~/models --max-model-memory 32GB
# Process-level memory limit (default: auto = RAM - 8GB)
omlx serve --model-dir ~/models --max-process-memory 80%
# Enable SSD cache for KV blocks
omlx serve --model-dir ~/models --paged-ssd-cache-dir ~/.omlx/cache
# Set in-memory hot cache size
omlx serve --model-dir ~/models --hot-cache-max-size 20%
# Adjust batch sizes
omlx serve --model-dir ~/models --prefill-batch-size 8 --completion-batch-size 32
# With MCP tools
omlx serve --model-dir ~/models --mcp-config mcp.json
# HuggingFace mirror endpoint (for restricted regions)
omlx serve --model-dir ~/models --hf-endpoint https://hf-mirror.com
# API key authentication
omlx serve --model-dir ~/models --api-key your-secret-key
# Localhost-only: skip verification via admin panel global settingsAll settings can also be configured from the web admin panel at /admin. Settings are persisted to ~/.omlx/settings.json, and CLI flags take precedence.
Architecture
FastAPI Server (OpenAI / Anthropic API)
│
├── EnginePool (multi-model, LRU eviction, TTL, manual load/unload)
│ ├── BatchedEngine (LLMs, continuous batching)
│ ├── VLMEngine (vision-language models)
│ ├── EmbeddingEngine
│ └── RerankerEngine
│
├── ProcessMemoryEnforcer (total memory limit, TTL checks)
│
├── Scheduler (FCFS, configurable batch sizes)
│ └── mlx-lm BatchGenerator
│
└── Cache Stack
├── PagedCacheManager (GPU, block-based, CoW, prefix sharing)
├── Hot Cache (in-memory tier, write-back)
└── PagedSSDCacheManager (SSD cold tier, safetensors format)
git clone https://github.com/jundot/omlx.git
cd omlx
pip install -e ".[dev]"
pytest -m "not slow"Requires Python 3.11+ and venvstacks (pip install venvstacks).
cd packaging
# Full build (venvstacks + app bundle + DMG)
python build.py
# Skip venvstacks (code changes only)
python build.py --skip-venv
# DMG only
python build.py --dmg-onlySee packaging/README.md for details on the app bundle structure and layer configuration.
Contributions are welcome! See Contributing Guide for details.
- Bug fixes and improvements
- Performance optimizations
- Documentation improvements
- MLX and mlx-lm by Apple
- mlx-vlm - Vision-language model inference on Apple Silicon
- vllm-mlx - oMLX started from vllm-mlx v0.1.0 and evolved significantly with multi-model serving, tiered KV caching, VLM with full paged cache support, an admin panel, and a macOS menu bar app
- venvstacks - Portable Python environment layering for the macOS app bundle
- mlx-embeddings - Embedding model support for Apple Silicon
- llm-compressor - Reference AWQ implementation for MoE models, used as design reference for oQ weight equalization