NV-Reason-CT ONNX for the browser
ONNX conversion of nvidia/NV-Reason-CT for ONNX Runtime Web with WebGPU. These files are used by https://jarrelscy.github.io/nv-reason-ct-web/ (source: https://github.com/jarrelscy/nv-reason-ct-web), which runs the model entirely in the browser.
Research use only. Not a medical device. Not validated or approved for clinical use, and must not be used for diagnosis or treatment decisions.
Files
| path | contents | size |
|---|---|---|
vision.onnx, vision.data0 |
Primus 3D ViT and merger, int8 weights. Input a 192³ crop in [-1, 1]; output 13,824 tokens of width 2560 | 184 MB |
int8/decoder.* |
Qwen3.5-4B hybrid decoder, int8 weights (round-to-nearest, block 32) | 4.8 GB |
int4/decoder.* |
Same decoder, int4 weights (GPTQ, block 32, act order) | 2.8 GB |
tokenizer.json, tokenizer_config.json |
tokenizer from the original model | 20 MB |
Each *.json manifest lists the ONNX graph, its external data chunks (split to fit browser buffer limits) and the decoder state names. The decoders use fp32 activations and MatMulNBits weights, and the lm_head is pruned to the last token. Token embeddings are tied to lm_head, so they are not stored separately; the web app dequantises embedding rows from the lm_head weights (described under embed in decoder.json). Image tokens use 3D MRoPE positions as in the original model; see lib/nvreason.js in the source repo.
Accuracy
Teacher-forced agreement with the fp32 model on the reference reports of two test cases (top-1 token agreement, mean KL of the next-token distribution):
| build | case 1 | case 2 |
|---|---|---|
| int8 | 99.1%, KL 0.0002 | 99.2%, KL 0.0004 |
| int4 GPTQ | 97.8%, KL 0.020 | 97.7%, KL 0.066 |
With greedy decoding the int8 build produces the same report as the fp32 PyTorch model on the test case. The int4 build writes well-formed reports but more often drops or changes findings.
Licence
NV-Reason-CT is by NVIDIA and licensed under the OpenMDW License 1.1 (LICENSE). These files are derived from it and distributed under the same licence.