MTPLX

MTPLX is a native Mac application for running local large language models that uses native MTP speculative decoding to achieve faster generation speeds.

MTPLX is a native Mac application designed to run local large language models on Apple Silicon hardware. It uses native MTP speculative decoding to increase generation speed, achieving up to double the speed of standard decoding without sacrificing output quality. The application supports various models, including Qwen 3.8 Flash Next, Qwen 3.8 27B, Ternary Bonsai 2 27B, and MiMo V2.6 Qwen 9B. Developers and users can utilize the app as an OpenAI and Anthropic compatible local server, allowing integration with existing tools. It includes features for chatting, running coding agents, and a forge tool to convert Hugging Face models to MLX format for speed testing on local machines.

MTPLX at a glance

Best for: Software engineer
Runs on: macOS, CLI
Pricing: Free and open source
Released: May 2026
Category: AI coding tool
Website: mtplx.com

Key features

Native MTP speculative decoding for faster generation
OpenAI and Anthropic compatible local server
Model conversion from Hugging Face to MLX
Performance benchmarking tools for Apple Silicon
Support for coding agents and local chat interfaces
Offline execution on macOS

Use it for

Running local LLMs for coding tasks
Hosting OpenAI-compatible local API servers
Testing performance of large models on Apple hardware
Converting and optimizing Hugging Face models for local execution

Alternatives to MTPLX

oMLX — It uses paged SSD KV caching to reduce agent TTFT to under 5 seconds.
MLX LM — Runs LLMs and fine-tuning natively on Apple silicon hardware.
Rapid-MLX — Run 250 MLX-native models locally on Apple Silicon with high inference speeds.
OptiQ — Runs entirely on Apple Silicon without PyTorch or cloud dependence.
Strata Mlx — Runs 125B MoE models on Apple Silicon by streaming experts directly from disk.
AirLLM — Runs massive 70B+ parameter models on small 4GB consumer GPUs.
LMDeploy — Accelerates inference with high-performance engines and quantization tools.
LangChain4j — Integrates LLMs and vector stores directly into Java applications.
Soup — Fine-tune and align models like Llama-3.1-8B on a 4 GB laptop GPU.
Strata — Runs large AI models locally on consumer gaming PCs without server reliance.
Mesh LLM — Run large local models across multiple machines through one OpenAI-compatible endpoint.
MLC LLM — Enables native optimization and deployment of AI models across platforms.

All MTPLX alternatives