Nvidia's open Nemotron 3.5 Super VL targets rivals' enterprise niches

Nvidia is giving away a 120B-parameter model and telling enterprises that, with some tuning, it can outperform the most expensive closed systems on their own documents. The catch sits in the pricing: the weights are free, but running them in production is not.

The company released Nemotron 3.5 Super VL, an open vision-language model built for PDFs, charts, tables, images and webpages, with instruction-following and tool use. Details were reported by VentureBeat. Weights are on Hugging Face, and the model is also available through OpenRouter, ModelScope and NVIDIA NIM. Nvidia offers a free, rate-limited hosted API for development.

The licence is OpenMDW, which permits commercial use, fine-tuning and redistribution. Nvidia is also releasing training recipes and datasets. Production deployment, however, requires NVIDIA AI Enterprise at $4,500 per GPU per year, which covers software only.

The headline benchmark claims come with a condition: they apply to post-trained, custom versions of Super VL, not the base model. Across five tasks, Nvidia reports gains of 9% to 44%. On medical claims accuracy, a tuned version rose from about 28% to about 72%, against about 38% for Claude Opus 5. On security attack resilience, it went from about 35% to about 76%, against about 66% for GPT-6.1 Sol. Nvidia has not released task definitions, sample sizes or testing protocols for these figures.

Nvidia also claims speeds up to 5x faster than a comparable open model, and about 625 tokens per second on four B200 GPUs. The model was trained on about 26 trillion tokens, with 3.8 million post-training samples and reinforcement learning across roughly 100 environments. On the Artificial Analysis index the base model scores about 26, trailing GLM-5.3-Flash at 42 and Qwen3.8-Flash-Next at 40.

Nvidia released two other models alongside it. Bolt 74B-A9B, with 74B total and 9B active parameters, targets local coding and research agents. Bolt E4B, a 4GB model, is due on 23 October.

Nvidia is not selling a general-purpose champion. It is selling a starting point. The base model's index score of 26 puts it well behind two rival open models, so the pitch rests on customisation: bring your own data, tune it, and beat a frontier closed model on your narrow task. That pitch pressures API-only vendors in regulated, document-heavy work such as claims processing and security, where a tuned model on owned hardware can be cheaper to control.

The economics favour Nvidia either way. Free weights widen the funnel, and the $4,500 per GPU per year software fee monetises every serious deployment, on top of the hardware itself. The open licence and published recipes lower the cost of switching in, while the enterprise subscription captures the value.

Buyers should treat the numbers cautiously. Without task definitions, sample sizes or protocols, the 9% to 44% gains cannot be checked, and the comparison pits tuned Nvidia models against untuned rivals. Watch for independent replications, and for how the pending $12.93 billion Hugging Face acquisition, expected to close in the first half of 2027, shapes where Nvidia's open models get distributed.