Pi on a local model.

Start the server and launch Pi. The model runs on your Mac, the sampler is the model's own, and since 2.0.2 (9 Jul 2026) Pi's tool turns restore from the warm prefix cache. Created by Youssof Altoukhi, who brought native MTP to the Mac in April 2026.

One command

Install MTPLX from the DMG or with Homebrew, then:

brew install youssofal/mtplx/mtplx
mtplx start pi

mtplx start pi loads the model, serves it on 127.0.0.1:8000, and launches Pi against the local server. The app's one-click Pi launch does the same thing. The app and the CLI share one server, so mtplx start attaches to a model the app already has loaded instead of loading a second copy.

Manual setup

Pi uses the OpenAI-compatible endpoint, so a manual setup needs the same three values as any OpenAI-compatible client:

The served id follows the loaded pack; /v1/models lists it. If you bind the server to a non-localhost address you must start it with --api-key, and the server accepts that key as Authorization: Bearer or X-API-Key.

curl http://127.0.0.1:8000/v1/models

The model's own sampling contract

MTPLX serves each pack at its own sampler. For Qwen 3.8 that is the official Qwen 3.8 sampling: temperature 1.0, top-p 0.95, top-k 20. Acceptance of drafted tokens is probability-ratio min(1, p/q) with residual (p - q)+ resampling (Leviathan–Chen), so the output follows the model's distribution at that sampler and at any other temperature. MTPLX has never shipped a greedy-only path. Activations stay bf16 (fp16 on M1 and M2) on every pass, and the speculative path runs the same model as the plain path.

A request may still set temperature, top_p, top_k, presence_penalty and frequency_penalty; the server honors them per request. Penalties default to 0, which is an exact no-op that preserves MTP exactness. Qwen's guidance is to leave them at 0 for coding and agent work.

Warm tool turns

Since 2.0.0 (6 Jul 2026) MTPLX checkpoints the attention KV cache plus the recurrent and conv GDN state at commit boundaries, with speculation on. 2.0.2 (9 Jul 2026) extended the warm prefix restore from OpenCode to Pi, Claude Code and Cline. A Pi tool round that appends a result and resends the transcript restores the prefix instead of prefilling it again: mid-session tool rounds restore warm in under 2 s, and a 100k-token session restores in about 2 s after a restart instead of a five-minute cold prefill. Every response reports usage.cached_tokens, so you can see what was served from cache.

MTPLX 2.11 (4 Sep 2026) removed the dead time between tool turns. Measured on a 43k-token agent session on Flash-Next against the live daemon (M5 Max, 128 GB): the follow-up after a file write re-prefills 20 tokens in 0.12 s instead of 3,535 tokens in 3.7 s, the wait before a warm turn's first token is 0.01 s instead of 5 to 7.5 s, and a forced tool_choice round in a 41k-token session takes 0.58 s instead of 85 s. The fixes are in the engine, so Pi's tool rounds ride the same path.

Reasoning effort

In 2.10.0 (29 Aug 2026) OpenCode and Pi began defaulting to the family's agent-lane reasoning effort, medium for Flash-Next, instead of xhigh. On the multi-file agent task in that release note, wall clock went from 150.2 s to 44.2 s with the same correct output; the gain came from the lower reasoning-effort default rather than from decode speed. The dial lives on the MTPLX side, in the app or through mtplx settings get and mtplx settings set.

Community launcher

KrossKinetic/pi-mtplx on GitHub is a community launcher that runs Pi against MTPLX. It is maintained outside the MTPLX repository; the built-in path is mtplx start pi.

Model ids

The served id follows the pack.

Details for each pack, including the quantization map and the measured speeds, are on the models pages.