minimlx chat·qwen38-27b-mtplx·tools: 7·bash: on
› what is minimlx?
thinking…
Short answer wanted. Say what it is, where it runs, and that it stays on the machine.
I run Gemma 4, Qwen 3.6 and Qwen 3.8 entirely on your own Mac, through Apple's MLX — chat, agentic coding, voice, and an Anthropic-compatible server for Claude Code. Nothing leaves the device.
68 tok · 57.0 tok/s · thought 84 chars
› minimlx --help
- chat
- Interactive REPL — streaming output, reasoning, and a local tool loop: read, write, edit, grep, glob, ls, bash.
- ask
- One-shot prompt. Bare text works too, so
minimlx "explain crc32"is enough. - code
- One-shot programming task against your working tree, with the same tools.
- conversation
- Voice: speak → transcribe → generate → speak back, all on-device.
- serve
- Anthropic-compatible HTTP server, so Claude Code can run against a local model.
- pull · ls · rm
- Model and cache management, with aliases for everything in the library.
› minimlx models — measured on an M5 Max, 128 GB
| alias | build | decode | prefill | peak |
|---|---|---|---|---|
| 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 |
300-token generations at --temp 0.7, decode rate from a cold model.
Sustained load roughly halves all of it — an M-series laptop throttles well
before the model does. The answer above streamed at the top row's rate.
› install
$ pip install "minimlx[mtplx] @ git+https://github.com/bravenewxyz/minimlx" # then $ minimlx chat $ minimlx code "add a --json flag" # or point Claude Code at your own hardware $ minimlx serve $ export ANTHROPIC_BASE_URL=http://127.0.0.1:1234
macOS on Apple Silicon, Python 3.11+. MIT licensed.
› the real thing
