Nvidia releases open-weight Nemotron 3.5 Super VL model

Summary

Nvidia has released Nemotron 3.5 Super VL, a 120-billion-parameter open-weight model. The company claims the model is five times faster than its closest competitor. Nvidia is also offering two smaller edge models (24 GB and 4 GB) targeted at local and on-device agent workloads.

Why it matters

This launch signals a visible move by Western labs like Nvidia, Mistral, and Reflection AI to close the open-weight model gap that Chinese labs such as DeepSeek and Moonshot have held for years. For enterprise customers, the ability to fine-tune on private data and run inference on their own infrastructure directly addresses data sovereignty and cost pressures, which are key factors accelerating the adoption of open-weight models. The launch also arrives at a time when the Hugging Face security incident is reshaping how buyers assess the risk balance between open and closed models.

Highlights

• According to Artificial Analysis, the 120-billion-parameter model is five times faster than the Ling 3.0 Flash (InclusionAI) and V4.1 Flash (DeepSeek) models.
• Post-training on a health insurance reimbursement task improved accuracy by 44%, surpassing Claude Opus 5 in this narrow domain.
• Two edge variants—Bolt 74B-A9B (24 GB) for local agents and Bolt E4B (4 GB) for on-device chat agents—aim for 'intelligence per gigabyte' efficiency.

Critical take

The speed and accuracy claims in the headline rely on a single benchmark source (Artificial Analysis) and a single narrow post-training example. These two data points cannot be generalized to the broad, weeks-long enterprise agent workflows Nvidia proposes. Furthermore, the article uses the Hugging Face breach to frame open-weight models as a security solution, yet a 120-billion-parameter weight file distributed publicly carries its own supply chain and data leakage risks, which the article fails to question.