Key points
- Nvidia launched Nemotron 3.5 Super VL and two Bolt edge models.
- The release follows a security incident at Hugging Face.
- Open models allow local post-training for enhanced data privacy.
Nvidia released the Nemotron 3.5 Super VL model alongside two new Bolt variants, joining a surge of open-weight AI tools from US laboratories. This launch occurs shortly after the Hugging Face hack, prompting developers to consider model integrity. The 120-billion-parameter Super VL targets agentic workflows, aiming for speed and accuracy in specialized tasks. According to Nvidia, the model delivers the fastest task completion in its class, helping enterprise agents handle high-volume operations.
The company claims Super VL is five times faster than competitors like InclusionAI’s Ling 3.0 Flash and DeepSeek’s V4.1 Flash, based on Artificial Analysis benchmarks. While the base model does not surpass proprietary frontier models alone, post-training yields significant gains. Nvidia cited a 44% improvement in medical insurance claims processing, outperforming Claude Opus 5. This allows businesses to tailor the model to specific needs while maintaining control over their infrastructure.
What we know
Nvidia introduced two additional edge models: Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB. These models operate locally and off-grid, offering high intelligence per gigabyte of memory. The 74B variant supports local agentic use, while the 4GB version handles fast conversational agents. Open models provide advantages over proprietary ones by enabling post-training on private data. This approach offers greater personalization and security for enterprises running models on their own hardware.
Defender's checklist
- Verify model hashes against official Nvidia repositories.
- Check network logs for unauthorized data exfiltration.
- Review access controls for AI development environments.
- Audit third-party dependencies for known vulnerabilities.
What analysts watch for
Teams usually monitor for unusual outbound traffic from AI training nodes, which may indicate data theft or command-and-control communication. Analysts examine model weights for embedded malicious code or backdoors that could activate during inference. They check whether training datasets contain poisoned samples designed to alter model behavior. Security teams also review access logs for unauthorized modifications to model repositories. Regular integrity checks of stored models help detect tampering after deployment.
What to do and how to stay safe: Nvidia
- Validate checksums of downloaded models against vendor-provided values.
- Isolate AI training environments from sensitive corporate networks.
- Monitor for unexpected resource consumption on GPU servers.
- Apply least-privilege access to model repositories and training data.
Step-by-step guide: Account Takeover: Definition, Vectors, and Defense Strategies
General security guidance from the Malware Chronicle newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How does Nemotron 3.5 Super VL compare to other models?
It is five times faster than competitors like Ling 3.0 Flash and V4.1 Flash, according to Artificial Analysis benchmarks.
What are the new Bolt models used for?
The 74B model supports local agentic use, while the 4GB model handles fast conversational agents.
Why are open models gaining traction now?
US labs are releasing more open models after the Hugging Face hack, challenging Chinese dominance in the sector.
