youssofal/MTPLX
Some data may be outdated — see provenance labels.
The fastest way to run Qwen 3.8 Flash Next, Qwen 3.8 27B and Ternary Bonsai 2 27B on a Mac: 125 tok/s in OpenCode on an M5 Max, and a 27B model on 16 GB Macs. Native MTP speculative decoding on Apple Silicon, exact at any temperature. OpenAI and Anthropic compatible local server. observed · 2026-10-04
Health v2 · maintenance only
81/100
- Activity 98
- Release rhythm 98
- Longevity 12
Flags: young
How is this computed?
round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10) — computed 2026-10-21. Adoption (stars, forks) is never an input.
- gap_med: 1
- age_days: 172
- days_rel: 18
- days_push: 17
- n_releases_24m: 52
Adoption not part of the score
2511 stars · 197 forks observed · 2026-10-04
What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded
MTPLX is a native macOS app and CLI that runs local LLMs on Apple Silicon roughly twice as fast using native multi-token prediction (MTP) speculative decoding, with exact rejection sampling that preserves output distribution. It ships an OpenAI/Anthropic-compatible local server, a built-in chat UI, and a model forge that converts Hugging Face models to MLX.
Use cases
- run local llms faster on my mac
- serve an openai-compatible local model endpoint on apple silicon
- speed up qwen inference with speculative decoding
- chat with local models offline on m-series macs
- convert a hugging face model to mlx and measure speedup
- power coding agents like opencode with a local model
- run llm inference without an external draft model
When to choose
- you have an Apple Silicon Mac (M1+, macOS 14+) and want faster local LLM inference with identical output
- you use MTP-capable models like Qwen 3.5/3.6/3.8 and want native speculative decoding without a second draft model
- you need an OpenAI- or Anthropic-compatible local server for coding agents and tools
- you want a one-click Mac app experience including model download and hardware-based recommendations
When to avoid
- you run on Linux, Windows, or non-Apple hardware
- you need models without MTP heads or broad model-family support
- you need multi-GPU or datacenter-scale serving rather than single-machine local inference
- you require a runtime with a long track record or large maintainer team
Facets
application · maturity active
llm-inference cli chatbot large-language-models machine-learning self-hosted developer-tools cli speculative-decoding multi-token-prediction mlx apple-silicon openai-compatible anthropic-compatible local-llm qwen metal inference-engine macos apple-ecosystem
5 sources
- readme: https://github.com/youssofal/MTPLX · fetched 2026-10-04 · 9e48cb2500d7
- homepage: https://mtplx.com · fetched 2026-08-29 · 12524fe1f9c4
- registry_pypi: https://pypi.org/pypi/mtplx/json · fetched 2026-08-29 · 77d01c735c27
- site_page: https://mtplx.com/faq · fetched 2026-08-29 · e8935b0f4d53
- site_page: https://mtplx.com/releases · fetched 2026-08-29 · 74e246e30bb8
Member repositories
| Repository | Role | Health v2 |
|---|---|---|
| youssofal/MTPLX | main | 81 |
For agents
Data as of 2026-10-19T08:05:51.212177+00:00 · Report a problem