Nvidia released Nemotron 3.5 Super VL, a 120B open-weight agentic model

Nvidia released three Nemotron 3.5 models with open weights, open datasets and open training recipes. The headline model, Super VL, has 120 billion parameters, Nvidia claims it reaches frontier-level intelligence at about one fifth of the parameters of the company's largest reasoning model, and ships for agent work inside the enterprise. Two smaller models accompany it: a 74B edge model and a 4B conversational model.

What is in the lineup

Super VL is the enterprise model, out at the announcement. Nvidia says it ranks first on PinchBench among open models in its class and second on the BrowseComp leaderboard, and that it delivers up to five times the throughput of comparable models at this size. The company has not published its own throughput number against a specific competitor, and the ranks are tied to the benchmarks the vendor is quoting.

Below Super VL sits Nemotron 3.5 Bolt 74B-A9B 24GB, released at the same time, built for local agentic tasks on an AI PC. Nvidia describes it as balancing speed and reasoning for multi-step work, tool calling and web browsing on an RTX Spark system. The third model, Nemotron 3.5 Bolt E4B, is a 4-gigabyte conversational model for the same local setup and arrives later this month.

All three models are published through Hugging Face, ModelScope, Open Router and NVIDIA NIM, with the datasets and recipes made available alongside, so a team can post-train and redeploy on its own infrastructure.

Who Nvidia names as a customer

The post names several companies as early users: Scale AI, Skydance, EY, Hebbia, AI21, CrowdStrike, Salesforce and Perplexity. Reporting in TechCrunch cites a similar list that also includes Ramp. Nvidia frames the choice around on-premises control: a team can post-train Super VL on its own data and keep inference inside its own infrastructure. The company's own paragraph on competition says Super VL completes agent tasks faster than models from Anthropic and OpenAI on document intelligence, financial accounting and security resilience. That is Nvidia's claim against named competitors, so it is a vendor comparison rather than an independent result and should be read that way.

What is new for a working engineer

The news is a frontier-grade open-weight model, with the datasets and recipes attached, in a size that fits inside a serious data centre budget. For a team that has been running an agent through a hosted API because the alternative open-weight model was too small or too limited, Super VL is a different choice. The 74B Bolt model moves the same question down to a workstation: Nvidia is positioning it on a laptop or a desktop with an RTX Spark chip, running the agent locally, calling tools locally, and keeping the request data on the machine.

Three facts are worth checking before committing to Super VL. Its benchmark ranks are open-model ranks against named leaderboards, which say nothing about how it will perform on your workload. The five-times throughput number is at a model-class level, not against a specific competitor model with a specific setup. And the Anthropic and OpenAI speed comparison is Nvidia's own, so a team that cares about that number should expect to measure it themselves.

For a team that was already deploying an open model and hitting the limits of what it could do, this release has more capability in the same open-weight shape, with the training materials in the open so the model can be modified rather than just used.

Source

Nemotron 3.5 Super VL: Scaling Agentic AI for the Enterprise and PC, Nvidia. Nvidia launches new open models to run agents in the enterprise and on PCs, TechCrunch.