Credits¶
This repository is glue: a server, launchers, an agent and installers. The hard parts underneath were built by other people.
Upstream software¶
| Project | By | Used for |
|---|---|---|
| Claude Code | Anthropic | The coding agent you run. Not modified; this project only points it at a local server. |
| MLX | Apple | Array framework on Apple Silicon's GPU and unified memory |
| mlx-lm | Apple | Model loading, chat templates, KV caches and stream_generate, used by the server and the Native Engine |
| mlx-vlm | Prince Canuma (Blaizzy) and contributors | Runs the vision builds (Muse-Glimmer) mentioned in the docs; not used by the server |
| mlx-dspark | ARahim3 | Speculative decoding engine used for the Qwen 3.8 DFlash 2 measurements |
| Hugging Face Hub | Hugging Face | Where models are downloaded from |
| OpenRouter | OpenRouter | Free cloud models offered by keep going |
| MkDocs and Material for MkDocs | Tom Christie and contributors; Martin Donath | This documentation site |
The repo's own credits name the people behind MLX and mlx-lm by commit count: Awni Hannun, Cheng, Angelos Katharopoulos, Ayaan Gazali, Jagrit Digani, Gökdeniz Gülmez, Anchen and everyone else who has contributed.
Models¶
Original releases:
| Model | By |
|---|---|
| Gemma 4 (31B, 12B, E4B) | Google DeepMind |
| Qwen 3.8 27B and Qwen 2.5 / Qwen3-Coder | Alibaba Qwen team |
| Hermes 4 14B | Nous Research |
| Muse Glimmer 30B | Meta |
| Nemotron 3 Nano Omni | NVIDIA |
| DFlash 2 drafter | Inco AI / Z Lab |
MLX conversions installed by the scripts come from mlx-community, lmstudio-community and the maintainer's own uploads at huggingface.co/divinetribe. The abliterated builds start from abliterations by huihui-ai, null-space, OpenYourMind and Babsie. The technique is explained by Maxime Labonne in Abliteration explained. Each model's own license applies.
Contributors to this repository¶
| Who | What they fixed |
|---|---|
| @0xshugo | Client disconnects handled, retries skipped when there are no tools (#4); diagnosed the wrong-model problem behind the launchers' model-aware restart (#5) |
| @asdmoment | Gemma inference crash: auto-disable KV quantization (#7) |
| @kulveersingh | ArraysCache has no attribute offset (#10) |
| @tripathiprateek | uninstall.sh (#23) |
| @tadrianonet | 16 GB Mac support, ChatML stop markers, the <tools> parser, server-sent events for tools and the offline launcher settings (#32) |
| @kevbarns | Gemma 4 thinking suppression and slimmer tool descriptions (#33) |
| @KaoCSC | Real end-of-sequence stop and empty env values (#41); bare JSON tool calls (#43); the prompt-cache and KV-quantization reports (#42, #46) |
| @zwolf25 | Slash commands in code mode and array-shaped fenced tool calls, with a parser test (#56) |
The server, launchers, Native Engine, keep going and install scripts are maintained by Matt Macosko. Bug reports and pull requests are welcome; see CONTRIBUTING.md.