📰 Key Takeaways

Nvidia rolled out its next-gen open-weight models on Thursday, built for “always-on AI agents” and edge computing, split across two tiers. The enterprise flagship is Nemotron 3.5 Super VL, a 120-billion-parameter model that’s smaller than Nemotron 3 Ultra but, according to Nvidia, matches it on instruction-following and tool-use tasks. On the other end are two lightweight Bolt models, built for the new AI PCs running Nvidia’s RTX Spark chip and for DGX Spark workstations, targeting consumers, engineers, and enterprise knowledge workers. They’re rolling out through partners like Microsoft and Dell.

Nvidia says that after agent-specific post-training, Super VL outperforms competing models from Anthropic and OpenAI (including Opus 5, GPT 6.1 Sol, Luna, and Astra) on tasks like document intelligence, financial accounting, scientific experimentation, medical insurance claims, and cybersecurity defense. The company also listed several large enterprise customers already using Super VL, including Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.

There’s a bigger strategic play behind Nvidia’s open model arsenal: as major AI labs start shifting toward other chipmakers like AWS Trainium and Cerebras — or even building their own chips — Nvidia is doubling down on the trend of enterprises running their own data centers with open models. The pitch: companies can post-train Super VL on their own data, deploy it on their own infrastructure, and keep full control over the training output. At the same time, Nvidia’s push on small PC-side models is trying to drive broader adoption of consumer AI assistants, though the article notes that as more AI assistants want deeper access to users’ home, health, and financial data, consumer wariness remains a real concern.

💬 JudyAI Lab Take

Nvidia splitting its open-weight lineup into a big enterprise model and lightweight PC models sends a clear signal: the chipmaker isn’t content just selling hardware anymore — it wants a say in training and deployment too.

For AI builders, what’s actually worth paying attention to isn’t the parameter-count comparison — it’s the option Nvidia is emphasizing: post-train on your own data, deploy on your own infrastructure. As major AI labs start spreading their compute across different chip vendors, or even building their own silicon, open-weight models plus self-hosted deployment are becoming a real third path alongside subscription APIs — one where companies keep full control over their training output. But getting lightweight consumer models to go mainstream is a different story — users’ wariness about AI assistants digging into their home, health, and financial privacy won’t just go away because the model got smaller.

If you’re evaluating an LLM deployment strategy, this is a good moment to re-run the build-vs-subscribe math on cost and control.

📅 Original Article Info

🔗 Further Reading