Text backbone: NemotronH hybrid (Mamba2 + MoE + attention), 30B total / ~3B active, 128-expert top-6, hidden 2688, 52 layers, 262k context.
Vision tower: C-RADIOv4-H ViT (InternVL-style pixel-shuffle head → mlp1 projector).
Audio tower: Parakeet Conformer (learned-filterbank frontend + BatchNorm conv module + Transformer-XL relative-position attention → sound_projection).

Conversion

Quantized with a first-party streaming quantizer (group_size 64, 4-bit affine) — the stock mlx_lm / mlx_vlm convert tools do not yet support the NemotronH_Nano_Omni_Reasoning_V3 multimodal wrapper. The LLM-backbone tensor layout is byte-identical to the text sibling mlx-community/NVIDIA-Nemotron-3-Nano-30B-A3B-4bit; the vision (vision_model.*) and audio (sound_encoder.*) towers are kept in bf16. ~19 GB on disk, ~20 GB working set (Apple-silicon Mac tier).

Use

The text backbone loads with standard MLX nemotron_h tooling. The vision and audio towers require a multimodal runtime that implements the C-RADIO ViT-H and Parakeet Conformer forward passes (e.g. the Evorix on-device engine).

License

Derivative of nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 under the NVIDIA Open Model License; original model © NVIDIA.