NVIDIA Nemotron 3.5 Super VL 120B A12B NVFP4 released

This is the deployment version of NVIDIA's Nemotron 3.5 Super VL, a 120-billion-parameter vision-language model that reads text, images and short videos. It is a mixture-of-experts model, so only about 12 billion parameters do work for any one token. The full-precision release stores weights in 16-bit BF16 and is listed for four H100 GPUs or two GB200s; this NVFP4 release stores most of them in NVFP4, a 4-bit format, and the card lists one DGX Spark, one H200, two H100s or one GB200 as deployment targets.

To recover accuracy after quantizing, NVIDIA applied quantization-aware distillation, further training in which the 4-bit model learns to match the full-precision model on the fine-tuning data. Some parts were left at higher precision: the vision encoder, the connector that feeds images into the language model, the Mamba convolutions and the embeddings stay in BF16, and the router that picks experts stays in 32-bit. On Blackwell GPUs inference runs with both weights and activations in 4 bits (W4A4); on other GPUs only the weights are 4-bit (W4A16).

The card's side-by-side table, run with the same settings for both checkpoints, shows most scores within about a point: SWE-bench Verified is 71.10 for BF16 and 71.07 for NVFP4, and GPQA-Diamond slightly favors NVFP4 at 89.90. A few tasks fall further, including MMLU-ProX (76.91 to 73.76), the WebVoyager browsing test (77.72 to 66.31) and DeepSWE (24.65 to 18.2). These are NVIDIA's own evaluations. For DGX Spark, the card pairs this checkpoint with a separate LiLiCorr drafter to speed up generation.