ETeissonniere/Nemotron-3-Nano-4B-CoreAI

NVIDIA Nemotron 3 Nano 4B — Core AI INT8

This is a native Apple Core AI conversion of nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 at immutable revision dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f. It is the Fast model used by the Qwen Core AI macOS exploration app.

The bundle uses the model's native hybrid architecture: Mamba-2 mixer layers, grouped-query attention, standalone MLP layers, recurrent convolution/SSM state, and attention KV state. It runs through Core AI; MLX and llama.cpp are not runtime dependencies.

Conversion

• Core AI profile: GPU-pipelined single-token decode
• Compression: selective INT8, block size 32
• Body weights: clipped symmetric quantization
• Untied vocabulary head: absolute-maximum symmetric quantization
• Embeddings, convolutions, RMSNorm, and SDPA composites: not quantized
• Operational context: 4,096 tokens
• Compiled payload: 4,609,042,537 bytes
• Payload SHA-256: d4967f627d20274ba8a06e8318f9f82e289b05c1c7e7f3b63b4c597cb35d0970

On an Apple M4 Pro with 48 GB unified memory, macOS 27 beta, and Xcode 27 beta, the packaged canary measured approximately 37 tokens/s prefill, 50 tokens/s decode, and 0.67 seconds to first token for a 25-token prompt. Performance depends on hardware, OS, prompt length, and thermal state.

Layout

The Core AI language bundle is at:

gpu-pipelined/nemotron_3_nano_4b_decode_int8hu

It contains the compiled .aimodel, embedded tokenizer, chat template, and Core AI language-model metadata. SHA256SUMS covers the published files, and PROVENANCE.json records the source and conversion identity.

License and attribution

The converted weights are distributed under the NVIDIA Nemotron Open Model License. See LICENSE and NOTICE before redistribution.

Licensed by NVIDIA Corporation under the NVIDIA Nemotron Model License.

This conversion is an independent project. It is not produced, endorsed, or certified by NVIDIA.