mtplx

The fastest way to run Qwen 3.8 Flash Next and Qwen 3.8 27B on a Mac: 125 tok/s in OpenCode on an M5 Max. Native MTP speculative decoding on Apple Silicon, exact at any temperature. OpenAI and Anthropic compatible local server.

Run local MLX models with Pi — zero config, automatic model discovery, and live token-speed monitoring.

One-click uncensored AI agent server for Apple Silicon Macs. Two engines - llama.cpp (GGUF) and MLX (MTPLX) - OpenAI-compatible on your LAN. Gemma 4 and Qwen. Up to 141 t/s decode.

One-click uncensored AI agent server for Apple Silicon Macs. MLX + MTPLX speculative decoding, OpenAI-compatible on your LAN. Up to 79.4 t/s decode, MTP speculative decoding on.

CLI to deploy and control a fleet of LLMs on local machine / LAN / Tailscale or AWS. Spin up your agent loop.

Double Qwen 3.8 27B inference speed on Apple Silicon with a one-click local coding agent

Beautiful, zero-dependency realtime dashboard + live activity log for a local MTPLX inference server — reads the /metrics endpoint, no build step.

Run an isolated Claude Science app copy through local or OpenAI-compatible model backends.

Native macOS menu-bar controller for local MTPLX model daemons.

Recover the native MTP predictor missing from the 8-bit MLX Qwen3.8-27B-Uncensored package, build a BF16 sidecar, and reproduce a 15.59 → 48.75 tok/s controlled M4 Max result with MTPLX.

Local OpenAI/Anthropic-compatible LLM gateway that starts, stops and switches local runtimes (MTPLX, LM Studio, oMLX, Ollama) behind one stable endpoint for your coding agents.

A heads-up display (HUD) with per-turn telemetry for both local inference engines and remote models. Look for it in the sidebar.