Nvidia releases open Nemotron 3.5 Super VL, 120B parameters

What happened: Nvidia released Nemotron 3.5 Super VL, an open 120B-parameter model for PDFs, charts, tables, images and webpages. It also handles instruction-following and tool use, according to VentureBeat.

The numbers: Nvidia says post-trained custom versions beat Claude Opus 5 and GPT-6.1 Sol on specialised tasks, across five tasks with gains of 9% to 44%. On medical claims accuracy, a custom version rose from about 28% to about 72%, against about 38% for Claude Opus 5. On security attack resilience, it rose from about 35% to about 76%, against about 66% for GPT-6.1 Sol.

The catch: These claims cover customised versions, not the base Super VL. Nvidia has not released task definitions, sample sizes or testing protocols for the benchmarks.

The context: On Artificial Analysis' index, Super VL scores about 26. That trails GLM-5.3-Flash (42) and Qwen3.8-Flash-Next (40).

Speed and training: Nvidia claims up to 5x the speed of a comparable open model, at about 625 tokens a second on four B200s. Training used about 26 trillion tokens, 3.8 million post-training samples and reinforcement learning across about 100 environments.

Who's affected: Weights are on Hugging Face, OpenRouter, ModelScope and NVIDIA NIM. Nvidia offers a free, rate-limited hosted API for development. The OpenMDW licence allows commercial use, fine-tuning and redistribution. Production use needs NVIDIA AI Enterprise at $4,500 per GPU per year, for the software only.

Why it matters: Nvidia is releasing training recipes and datasets alongside the weights. That lets teams build their own specialised versions, which is where its benchmark claims apply.

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