minimlx — local LLMs on Apple Silicon

What it does

Interactive chat

Streaming output, reasoning display, and a local tool loop — read, write, edit, grep, glob, ls, and bash.

Agentic coding

minimlx code runs one-shot programming tasks against your working tree with the same tools.

Voice conversation

Speak, transcribe, generate, speak back. Fully local, no cloud round trip.

Backend for Claude Code

minimlx serve speaks the Anthropic API, so Claude Code can run against a local model.

Speculative decoding

DFlash block-diffusion drafts, Google MTP drafters, and MTPLX builds that carry their own prediction head.

Prompt & KV caching

Cross-request prefix reuse, model aliases, and conversation logging out of the box.

Measured on an M5 Max

Model	Build	Decode	Prefill	Peak RAM
`qwen38-27b-mtplx`	27B, 4/8-bit + MTP head	57 tok/s	352 tok/s	21.0 GB
`qwen38-9b-4bit`	9B, 4-bit	98 tok/s	358 tok/s	5.3 GB
`qwen38-9b`	9B, 8-bit	58 tok/s	276 tok/s	9.7 GB
`qwen38-27b-4bit`	27B, 4-bit	32 tok/s	234 tok/s	15.5 GB
`qwen38-27b`	27B, 8-bit	16 tok/s	159 tok/s	28.9 GB

MacBook Pro, Apple M5 Max, 128 GB unified memory. 300-token generations at --temp 0.7, decode rate from a cold model. Sustained load roughly halves these — an M-series laptop throttles well before the model does.

Getting started

# install (or clone first, if you want to hack on it)
$ pip install "minimlx[mtplx] @ git+https://github.com/bravenewxyz/minimlx"

# talk to it
$ minimlx chat
$ minimlx ask "explain crc32"
$ minimlx code "add a --json flag"

# point Claude Code at a local model
$ minimlx serve
$ export ANTHROPIC_BASE_URL=http://127.0.0.1:1234

Requires macOS on Apple Silicon and Python 3.11+.

Local models,
at laptop speed.