Nvidia's open models, from server to laptop

Nvidia has released a new set of open-weight models for AI agents, from data-centre servers down to laptops. In a blog post dated Thursday, 15 October, Kari Briski, Nvidia's vice president for open models and software, announced Nemotron 3.5 Super VL, a 120-billion-parameter model for enterprise agents, and two smaller "Bolt" models for local machines. Super VL and the larger Bolt model are available now; the smaller Bolt is due on Friday, 23 October.

Nvidia says Super VL matches the accuracy of its largest reasoning model on agent tasks such as following instructions, using tools, running multistep business workflows and understanding images, at roughly one-fifth of the parameter count. It calls it "the fastest open agentic model in its class" and says it is up to five times faster than the next best. Those are Nvidia's figures. The model reads text, web pages, charts and PDFs, and ships with open weights, open datasets and training recipes, the company says, and is available on Hugging Face and OpenRouter, among others.

The larger edge model, Nemotron 3.5 Bolt 74B-A9B 24GB, has 74 billion parameters, of which about 9 billion are active at a time, and uses 2-bit quantisation to fit in 24GB of memory. Nvidia says it is built for its RTX Spark laptops and also runs on GeForce RTX PCs and DGX Spark and DGX Station workstations, and that it ranks first among open models of its class on the PinchBench benchmark. Users can hand it tasks "from researching the web to writing to debugging code — all without sending those workloads to the cloud", Briski wrote. Bolt E4B 4GB is meant for voice assistants and other conversational agents.

Nvidia named a long list of companies adapting Super VL. It says Scale AI post-trained the model for agents at Skydance and EY, and that Ramp, Lila Sciences and R1 are among others customising it; Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys are integrating it into their software. "Unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and oversee what they build," Briski wrote.

TechCrunch, which wrote about the models on Monday, said Nvidia claims Super VL completes tasks faster than competing models from Anthropic and OpenAI when it is post-trained on specific agentic tasks. That comparison does not appear in Nvidia's announcement or in its technical blog post, and no figures for it have been published. Nvidia's announcement also does not name a licence beyond describing the weights as open.

TechCrunch's reading of the strategy is that the closed-model labs that are both Nvidia's customers and its investments are buying more chips from other suppliers, such as Amazon's Trainium and Cerebras, or designing their own, so Nvidia is courting companies that want to run open models on their own hardware. That is the publication's analysis, not Nvidia's. For laptops, the pitch is privacy: agents that work on a user's files without the data leaving the machine.

Komposite has covered the wider open-weight field, including Meta's release of an 816-billion-parameter model and two "open" models from Mistral and Reflection whose weights had not yet shipped.