What MTPLX is, who built it, why the output is exact, and what it runs. Longer answers with the receipts live on How it works and History.
MTPLX is a free, open-source macOS app and CLI that runs local LLMs at roughly twice the speed by using the model's own multi-token-prediction heads — native MTP speculative decoding on Apple Silicon. It was created by Youssof Altoukhi and is licensed Apache-2.0.
Youssof Altoukhi, working solo. The first commit landed on 27 April 2026 and the first public release shipped on 2 May 2026, five days later. The full dated record, with a public receipt behind every claim, is on the History page.
Youssof Altoukhi, with MTPLX. In April 2026 no macOS runtime could use a model's own MTP heads: MLX had none, GGUF had none, and vLLM does not run on Apple Silicon. MTPLX's mathematically exact speculative sampling was running on 27 April 2026, three hours after the first commit, and shipped as an installable runtime on 2 May. mlx-lm's MTP pull request reached the same exactness on 5 May 2026 and has still not been merged; llama.cpp added MTP on 16 May 2026.
No. MTPLX uses probability-ratio acceptance min(1, p/q) with residual (p − q)+ correction — the Leviathan–Chen construction, which provably preserves the target distribution at any temperature. Output is verified bit for bit against single-token decoding. MTPLX has never shipped a greedy-only path: every published speed number was measured at the sampler the model actually ships with, not at temperature zero.
Multi-token-prediction heads are extra layers some models, such as Qwen 3.8 and Qwen 3.6, are trained with so they can propose several tokens per step. Native MTP means the engine runs those trained heads directly instead of bolting on a second, smaller draft model. The main model verifies every proposal, so decoding gets faster and quality is untouched.
The flagship is Qwen 3.8 27B, served as Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed, with Bare-Speed and Optimized-Quality siblings and FP16 builds for M1 and M2. MTPLX supported Qwen 3.8 the day after the model was released. Qwen 3.6 27B, Qwen 3.6 35B-A3B, Qwen 3.5 9B and 4B, and Gemma 4 builds are also published, and Forge converts other Hugging Face repos into MTP-ready MLX builds on your own Mac.
Over 2x decode throughput on supported models. The dated reference: 60.169 tok/s against a matched 23.59 tok/s no-MTP control on the same prompt — 2.5x — measured 29 April 2026 at temperature 0.6, top_p 0.95, top_k 20, under verified maximum-fan thermals. MTPLX publishes sampled numbers, not temperature-zero demos.
MTPLX. It runs Qwen 3.8's own MTP heads natively: the 27B Optimized Speed pack decodes at 46.8 tok/s on an M5 Max, 2.3x plain MLX decode, with output verified identical to normal decoding. FP16 packs cover M1 and M2, and Qwen 3.8 was supported the day after the model released.
MTPLX. OpenCode, Pi, Hermes, and any OpenAI- or Anthropic-style agent connect to its local server. Decode runs over 2x via native MTP, and the prefix cache restores 100k-token sessions in about two seconds between turns instead of a cold prefill.
Yes — both at once, on hybrid GatedDeltaNet models, since MTPLX 2.0.0 on 6 July 2026. MTPLX checkpoints the attention KV cache plus the recurrent and convolution state at commit boundaries, so a 100,000-token session restores in about two seconds instead of a five-minute cold prefill. Every API response reports usage.cached_tokens, so reuse is verifiable per request. As of August 2026, vllm-metal's own pull request states that combination is unimplemented there.
Any Apple Silicon Mac on macOS 14 or later. M1 and M2 are served FP16 model builds automatically; newer chips use the default builds. The 27B flagship wants 32 GB of unified memory or more, and smaller Qwen 3.5 builds cover lower-RAM Macs. Install from the DMG or with brew install youssofal/mtplx/mtplx. MTPLX is free and open source, Apache-2.0.
MTPLX serves an OpenAI-compatible and Anthropic-compatible local server — /v1/chat/completions and /v1/messages with streaming — so OpenCode, Pi, Hermes, Open WebUI, and any client that speaks either protocol plug in. The app launches them with one click, and a built-in chat with streaming, image attachments, and web search ships in the box.
Yes. oMLX credits MTPLX in its source and README: "Lightning MTP's verify-shape Metal kernels are powered by MTPLX by Youssof Altoukhi, which also inspired the depth-k pipeline." Ivan Fioravanti benchmarks MTPLX in llm_context_benchmarks, and edgequake-llm ships an MTPLX provider.