Nvidia has added three open-weight models to its Nemotron family: Nemotron 3.5 Super VL, a 120-billion-parameter model for enterprise agents, and two smaller “Bolt” models designed to run on a user’s own PC or workstation. Two of the three are available now, and the smallest follows later this month, according to TechCrunch’s report on the launch. The release pairs a data-center pitch with a consumer-hardware pitch, and both serve the same strategic goal for the chipmaker: keeping agent workloads on Nvidia hardware.

What Nvidia released

The three models split cleanly by where they are meant to run.

  • Nemotron 3.5 Super VL has 120 billion parameters and is geared toward agents deployed inside companies. Nvidia says it delivers accuracy similar to the larger Nemotron 3 Ultra model on instruction-following and tool-use tasks while using fewer parameters.
  • Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model intended for local agentic work such as multistep reasoning, tool calling and web browsing. The name follows the convention used for mixture-of-experts models, where the second figure typically refers to the parameters active for each token; Nvidia’s materials, not ours, define the exact meaning here.
  • Nemotron 3.5 Bolt E4B is the smallest model and is aimed at conversational agents. It is due later this month.

Super VL and the 74B-A9B Bolt model are available now. The Bolt models are intended for the new AI PCs built on Nvidia’s RTX Spark chip and for workstations running the DGX Spark, hardware that partners such as Microsoft and Dell are bringing to market. We covered the first RTX Spark laptops in our earlier report.

Claims about performance, and who is using it

Nvidia says Super VL completes agentic tasks faster than competing models from Anthropic and OpenAI when it is post-trained on specific tasks. The comparison set named in TechCrunch’s report includes Opus 5, GPT 6.1 Sol, Luna and Astra. The tasks Nvidia cites are document intelligence, financial accounting, scientific experimentation, medical insurance claims and resistance to security attacks.

As with any vendor comparison, the key phrase is “when post-trained.” The result describes a model that has been specialized for a task, not a general-purpose head-to-head, and we have not seen independent measurements. Nvidia has listed companies it says are using Super VL: Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce. A customer list shows interest, not performance, but it does suggest the model is being tried in production-adjacent settings rather than only in a lab.

Why open weights matter for agents

Nvidia’s vice president of open models and software, Kari Briski, framed the release around ownership. In the launch post she wrote that organizations can post-train Super VL on their own data, deploy it on their own infrastructure and “own what they build,” in contrast to closed frontier models. She also said any organization wanting to run agents on Nvidia infrastructure now has an open, optimized model for its tier without having to choose between capability and deployability.

For companies handling sensitive documents, that matters. An open-weight model can sit inside a company’s own network, be fine-tuned on private data and be audited more closely than an API. The trade-off is that the company takes on the work of serving, securing and updating it. For background on that choice, see our comparison of open-weight and closed models.

The strategic context

Nemotron is not new. Nvidia already publishes a family of models covering multimodal AI, reasoning, coding, math and enterprise automation. What is changing is the market around it. The proprietary labs that buy Nvidia’s GPUs, and in some cases receive its investment, are increasingly looking at other suppliers, including AWS’s Trainium and Cerebras, and at chips they design themselves. TechCrunch describes Nvidia as leaning into the trend of enterprises wanting to run their own open models in their own data centers.

The Bolt models address a second bet: that people will want personal assistants running locally. Briski wrote that PC users can delegate everyday projects without sending those workloads to the cloud. TechCrunch notes the obvious tension. A useful assistant needs access to information about a person’s family, health and finances, and privacy-wary consumers may hesitate to grant that, even locally. Running on-device reduces data leaving the machine, but it does not by itself decide what the agent is allowed to do.

What to watch next

  • Independent benchmark results for Super VL against the proprietary models Nvidia names.
  • The release of Bolt E4B later this month and its real-world speed on RTX Spark PCs.
  • Licensing terms for each model, which determine how freely companies can fine-tune and redistribute them.
  • Whether major software vendors ship agents built on Nemotron rather than on closed APIs.
  • How memory limits on consumer PCs shape which Bolt model people can actually run.

Key takeaways

  • Nvidia released Nemotron 3.5 Super VL (120B) and Bolt 74B-A9B now, with Bolt E4B due later this month.
  • Super VL targets enterprise agents; the Bolt models target local agents on RTX Spark PCs and DGX Spark workstations.
  • Performance claims against Anthropic and OpenAI models come from Nvidia and apply to post-trained tasks.
  • The release supports Nvidia’s push to keep agent workloads on its hardware as rivals diversify chips.

Frequently asked questions

Are the Nemotron 3.5 models open source?

They are described as open-weight, meaning the trained parameters can be downloaded and post-trained. Open-weight is not always the same as open source, so check each model’s license before building on it.

Can I run the Bolt models on any laptop?

Nvidia designed them for RTX Spark PCs and DGX Spark workstations. Whether they run acceptably on other hardware depends on memory and the license, which we have not verified.

How is Super VL different from the Bolt models?

Super VL is a larger 120-billion-parameter model for enterprise agents running on infrastructure. The Bolt models are smaller and built for local use on a PC.