Credits¶
This repo is an MLX runtime for someone else's model. The pieces it stands on:
- NVIDIA built Nemotron-3-Nano-Omni-30B-A3B and released the weights and the PyTorch reference code openly. The vision, audio and processor code here is a port of that reference. The vision tower is NVIDIA's C-RADIOv4-H.
- yayr at mlx-community made the 4-bit MLX quantization this runs on.
- Apple's MLX team and the contributors of
nemotron_hto mlx-lm. The text backbone runs on their code unchanged. - Hugging Face transformers. The audio tower is a
port of its
ParakeetEncoder(v5.5.0), and the tokenizer is loaded through it. - @wacomctl672 converted the abliterated bf16 build to GGUF and documented running it on llama.cpp.
- mlx-vlm and mlx-audio by Blaizzy are the README's suggested alternatives for other model families.
- Anthropic's Claude appears as co-author on the commits in this repo's history.
The MLX runtime in src/, scripts/ and tests/ was written by
Matt Macosko (Nice Dreamz LLC). If you want local MLX models
behind Claude Code as well, see
claude-code-local.
License¶
Runtime code: MIT, see LICENSE.
Model weights are NVIDIA's and carry the
NVIDIA Open Model License.
This repo ships no weights.
This documentation site is built with MkDocs and Material for MkDocs.