Nvidia releases open Nemotron 3.5 Super VL vision-language model

Nvidia on 15 October released Nemotron 3.5 Super VL, an open vision-language model with 120 billion parameters, and claimed that customised versions outperform Claude Opus 5 and GPT-6.1 Sol on specialised tasks.

Nvidia said the model reads PDFs, charts, tables, images and webpages, and supports instruction-following and tool use, according to VentureBeat.

The benchmark claims cover five tasks and apply to post-trained custom versions, not the base model. In one test, accuracy on medical claims rose from about 28% to about 72%, against about 38% for Claude Opus 5. In another, security attack resilience rose from about 35% to about 76%, against about 66% for GPT-6.1 Sol. Gains across the tasks ranged from 9% to 44%.

Nvidia has not released task definitions, sample sizes or testing protocols for these claims. The claims are not a finding that the base Super VL beats those models in general.

Nvidia also said the model runs up to 5 times faster than a comparable open model, at about 625 tokens per second on four B200 GPUs. On the Artificial Analysis index, Super VL scores about 26. That trails GLM-5.3-Flash at 42 and Qwen3.8-Flash-Next at 40.

Nvidia said it trained the model on about 26 trillion tokens, with 3.8 million post-training samples and reinforcement learning across about 100 environments. It is releasing the training recipes and datasets.

The weights are on Hugging Face, OpenRouter, ModelScope and NVIDIA NIM. The licence is OpenMDW, which permits commercial use, fine-tuning and redistribution. Nvidia offers a free, rate-limited hosted API for development. Production use requires NVIDIA AI Enterprise at $4,500 per GPU per year, which covers software only.

Nvidia also released Bolt 74B-A9B, a model with 74 billion total and 9 billion active parameters, for local coding and research agents. A 4GB version, Bolt E4B, is due on 23 October. Separately, Nvidia's pending acquisition of Hugging Face, valued at about $12.93 billion, is expected to close in the first half of 2027.