Nvidia launches open-weight Nemotron 3.5 models for enterprise agents and local AI PCs

Nvidia has rolled out three new open-weight models in its Nemotron family, built for software agents that run inside company data centers as well as for smaller machines at the edge. According to TechCrunch, the release pairs a large enterprise-focused model with two compact "Bolt" models designed to run locally on personal computers and workstations.

A 120-billion-parameter flagship for enterprise agents

The largest model, Nemotron 3.5 Super VL, is a 120-billion-parameter system intended for enterprise agent workloads. TechCrunch reports that although it is smaller than Nvidia's Nemotron 3 Ultra, the company says it reaches comparable accuracy on instruction-following and tool-use tasks.

Nvidia also claims that when Super VL is post-trained for particular agentic jobs — document intelligence, financial accounting, scientific experimentation, medical insurance claims and security-attack resilience among them — it finishes work more quickly than rival models from Anthropic and OpenAI, which TechCrunch identifies as Opus 5, GPT 6.1 Sol, Luna and Astra. Benchmarks like these are a fixture of model launches, as TechCrunch notes, but Nvidia supported this one with a roster of companies it says already use Super VL, including Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce.

Two Bolt models for local machines

At the compact end, Nvidia introduced two Bolt models meant to run on the new generation of AI PCs built around its RTX Spark chip, and on workstations using the DGX Spark. Per the report, these machines — intended for consumers, developers and knowledge workers — are currently shipping through partners such as Microsoft and Dell.

The larger of the two, Nemotron 3.5 Bolt 74B-A9B 24GB, is a 74-billion-parameter model designed for local agentic work, keeping responses fast while still handling multi-step reasoning, tool calls and web browsing. The smallest, Bolt E4B, is built for conversational assistants.

TechCrunch reports that Super VL and Bolt 74B-A9B are available now, with Bolt E4B due later this month.

A bet on open models as labs look elsewhere

The release fits into a broader strategic shift. Nvidia's biggest AI customers, the frontier labs, are broadening their hardware options and exploring AWS's Trainium chips, Cerebras systems and their own in-house silicon, TechCrunch points out. In response, Nvidia is courting organizations that prefer to run and fine-tune their own models on their own infrastructure.

Kari Briski, Nvidia's vice president of open models and software, wrote in the announcement that any organization wanting to run agents on Nvidia infrastructure now has "an open, optimized model for their tier without having to choose between capability and deployability." She also emphasized that unlike closed frontier models, buyers can post-train Super VL on their own data, deploy it in-house and keep ownership of what they build.

The PC models target a second opportunity: consumer AI assistants. With privacy-minded users hesitant to give an assistant deep access to details about family, health and finances, Nvidia is betting that models running entirely on the device change the calculation. Briski wrote that local execution lets PC users delegate everyday projects "without sending those workloads to the cloud."

Why it matters

The launch shows Nvidia positioning itself as a full-stack supplier for the agentic AI era: chips, systems and now open models matched to each tier of hardware. For enterprises, an open-weight model that reportedly rivals larger systems on agentic accuracy offers something API-based frontier models cannot: control over data, fine-tuning and deployment. It also sharpens an awkward rivalry, since Nvidia is benchmarking its open models against proprietary labs that are simultaneously its customers and its investments. And if local, privacy-preserving assistants catch on, Nvidia would be supplying both the silicon and the model.