Sawfwair/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16-MLX-Native

Nemotron 3 Nano Omni 30B-A3B Reasoning BF16 — MLX Native

This is a lossless storage-layout conversion of NVIDIA's Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 checkpoint for native Swift/MLX inference on Apple Silicon.

The conversion does not quantize, round, train, or otherwise change tensor values. Every BF16 parameter payload is copied byte-for-byte and stored once:

• model-00001-of-00017.safetensors through model-00017-of-00017.safetensors contain the non-expert tensors under their upstream keys.
• experts-bf16.safetensors contains the routed expert tensors stacked into the physical layout consumed by the native MLX gather-matmul path.
• model.safetensors.index.json indexes only the non-expert shards.
• mere-run-native-checkpoint.json records the source revision and exact payload inventory.

The resulting checkpoint contains approximately 66.03 GB of model weights. It replaces the older workflow that retained the 66.03 GB upstream checkpoint and generated a second 58.75 GB expert cache at first launch.

Provenance

• Source repository: nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16
• Source revision: 24e67ea000b7c2837fc8f9488aa2008524fac8ba
• Native checkpoint format: mere-run-nemotron-omni-native-v1
• Precision: BF16
• Total weight payload: 66,031,270,520 bytes
• Expert payload: 58,749,616,128 bytes
• Non-expert payload: 7,281,654,392 bytes

Runtime

This repository uses a native expert layout and is not a drop-in Transformers checkpoint. It is consumed directly by the Nemotron Omni Swift/MLX runtime in mere.run.

The runtime supports text, image, audio, and video inputs with text output. It requires an Apple Silicon Mac with at least 112 GB unified memory; 128 GB is recommended.

License

The model remains governed by the NVIDIA Open Model Agreement. Review LICENSE.pdf, NOTICE, the upstream model card, and the accompanying safety, privacy, bias, and explainability documentation before use or redistribution.