Nvidia announced on Thursday a set of open-weight language models built for always-on agents and edge deployments. The new lineup includes a large enterprise-focused model and two smaller models designed to run locally on AI PCs and workstations.
Models and intended uses
-
Nemotron 3.5 Super VL: a 120-billion-parameter model aimed at enterprise agent workloads. Although it has fewer parameters than the Nemotron 3 Ultra, Nvidia says Super VL can match its accuracy on instruction-following and tool-use tasks.
-
Bolt family: two lightweight models intended to run locally on AI PCs equipped with the Nvidia RTX Spark chip and on workstations running DGX Spark. These machines are being introduced to the market through partners such as Microsoft and Dell and target consumers, developers, and enterprise knowledge workers.
Bolt model specifics:
- Nemotron 3.5 Bolt 74B-A9B 24GB: a 74-billion-parameter model for local agentic tasks that balances speed and reasoning to support multi-step reasoning, tool calling, and web browsing.
- Bolt E4B: the smallest Bolt variant, aimed at conversational agents.
Performance claims and enterprise adoption
Nvidia claims that when post-trained on agentic tasks, Super VL completes those tasks faster than certain competing models from other labs (the announcement named models such as Opus 5, GPT 6.1 Sol, Luna, and Astra). The benchmarked tasks include document intelligence, financial accounting, scientific experimentation support, medical insurance claims processing, and resilience to security attacks.
To support its performance claims, Nvidia listed several companies it says are using Super VL, including Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce.
Strategic rationale: openness and on-premises deployment
Nvidia emphasizes that open-weight models let organizations post-train on their own data, deploy on their own infrastructure, and retain ownership of what they build. Kari Briski, Nvidia’s vice president of Open Models and Software, wrote that “any organization that wants to run agents on Nvidia infrastructure now has an open, optimized model for their tier without having to choose between capability and deployability.”
At the same time, many proprietary frontier labs are exploring other chips and system vendors beyond Nvidia — for example AWS Trainium, Cerebras, or in-house silicon — prompting Nvidia to push for enterprises that prefer to run open models in their own data centers.
Local AI assistants and privacy
The Bolt models are also positioned to advance the early market for consumer AI assistants that run locally. Nvidia argues that devices which execute models and agents on the PC itself enable users to delegate everyday tasks without sending sensitive workloads to the cloud, addressing privacy concerns about granting remote services access to personal data.
Availability
Nvidia says Nemotron 3.5 Super VL and Bolt 74B-A9B are available now; Bolt E4B is expected later this month.
The announcement reinforces Nvidia’s dual strategy of serving enterprise customers that want open, on-premises models while also pursuing a market for local AI-enabled PCs and workstations that run models on Nvidia hardware.
