MTPLX

Project Overview

MTPLX is a native Mac app and CLI for Apple Silicon that doubles local LLM inference speed using Multi-Token Prediction (MTP). It leverages built-in MTP heads in modern models like Qwen 3.5/3.6/3.8 to achieve the same model output distribution with significantly faster speeds.

Key Features

• Multi-Token Prediction (MTP): The model drafts multiple tokens ahead and verifies each drafted block in a single batched forward pass
• Native Mac app: Dashboard with real-time tokens/sec, acceptance rate, cache state, and system pressure
• Built-in chat & agent server: File attachments, web search, one-click launch of OpenCode, Pi, Hermes, and more
• Auto-tune: Tests different draft depths during onboarding to pick the fastest configuration for your Mac
• Built-in AIME benchmark: Score models yourself with fully disclosed prompts

Target Audience

• Apple Silicon Mac users (M1 or newer)
• Developers who want to run LLMs efficiently on local hardware
• Users who追求 high-performance, low-resource local AI inference

Quick Start

1. Mac app (recommended): Download DMG from mtplx.com, drag to Applications
2. CLI:
   • Homebrew: brew install youssofal/mtplx/mtplx && mtplx start
   • pip: python3 -m pip install mtplx
3. The app checks hardware, recommends a model, downloads it, and sets up its own Python engine
4. Recommended for coding: Qwen 3.8 27B Optimized Speed (4-bit dynamic quant)
5. Requires macOS 14+ and Apple Silicon (M1 or newer). 16GB RAM runs 4B and 9B models comfortably

Core Advantages

• 2x+ speed improvement: 1.6x on M4 Mac mini, 2.24x on M5 Max
• No external drafter: Uses built-in MTP heads, no extra RAM consumption
• Exact sampling: Leviathan and Chen rejection sampling theorem ensures identical output distribution
• Native Mac experience: Beautiful dashboard and chat interface out of the box

System Requirements

• OS: macOS 14+
• Hardware: Apple Silicon (M1 or newer)
• RAM: 16GB runs 4B and 9B models comfortably; 32GB+ recommended for Qwen 3.8 Optimized Speed