Nvidia is now in the business of selling software that runs on chips it may never make. On October 15, the company released three open-weight AI models under the Nemotron 3.5 family, built for one specific job: running AI agents that work for hours, days, or weeks without a human watching. The launch, reported in detail by TechCrunch on October 19, puts Nvidia in direct competition with the frontier labs that are its biggest customers, and its biggest chip buyers.
The strategy is deliberate. As those labs shop for alternatives to Nvidia hardware, including AWS Trainium, Cerebras systems, and their own silicon, Nvidia is leaning into the one trend that favors it: enterprises that want to run open models in their own data centers instead of renting frontier models by the token.
Three models, three tiers: Super VL, Bolt 74B, Bolt E4B
The family covers the full range of where agents actually run, from a corporate data center down to a laptop on a desk.
| Model | Size | Built for | Availability |
|---|---|---|---|
| Nemotron 3.5 Super VL | 120 billion parameters (about 12 billion active per token) | Always-on enterprise agents: document review, web research, coding, cybersecurity triage, video analysis | Available now |
| Nemotron 3.5 Bolt 74B-A9B 24GB | 74 billion parameters | Local agentic tasks on AI PCs and workstations: multi-step reasoning, tool calling, web browsing | Available now |
| Nemotron 3.5 Bolt E4B | Smallest of the three | Conversational agents | Later this month |
Super VL is the headline model. It is a multimodal mixture-of-experts model that understands text, images, documents, and video in a single model, which matters for agents that need to read a spreadsheet, watch a clip, and browse the web in the same workflow. The two Bolt models are aimed at the new wave of AI PCs running Nvidia’s RTX Spark chip and DGX Spark workstations, hardware that partners like Microsoft and Dell are shipping now.
The performance claims Nvidia is willing to print
Nvidia’s numbers are aggressive, and the company knows they will be read as a direct challenge to its own customers.
Super VL, Nvidia says, matches the accuracy of its largest reasoning model with roughly one-fifth of the parameters, and delivers up to five times the throughput of comparable open models on agent workloads, citing Artificial Analysis leaderboard results. More pointedly, Nvidia claims Super VL completes tasks faster than Anthropic’s Opus 5 and OpenAI’s GPT-6.1 Sol, Luna, and Astra when the models are post-trained on specific agentic tasks: document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience.
The “when post-trained” qualifier does real work in that sentence. These are not raw out-of-the-box comparisons, they are comparisons after Super VL has been tuned on the exact task. That is also the entire sales pitch: because the weights are open, an enterprise can post-train Super VL on its own data and run it on its own infrastructure. As Kari Briski, Nvidia’s vice president of Open Models and Software, wrote in the announcement, “Unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build.”
Nvidia says Super VL is already in use at Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce. Take the customer list as a signal of who wants agent infrastructure they control, rather than as a benchmark result.
Why Nvidia is arming the enterprises
The interesting part of this launch is not the parameter counts, it is the timing. The same week these models arrived, the frontier labs kept diversifying their chip supply away from Nvidia. The labs are Nvidia’s biggest customers and also, increasingly, its competitors in the model business.
Open-weight enterprise models are Nvidia’s answer. Every company that runs agents on Super VL instead of a closed API is a company that needs Nvidia-class hardware, or at least the CUDA ecosystem, to run it well. It converts the model business from a threat into a hardware funnel. And it lets Nvidia compete in AI software without fighting the labs on their strongest ground, which is the closed frontier model itself.
The consumer play is the mirror image. The Bolt models are meant to push the nascent trend of local AI assistants on AI PCs. As Briski wrote, PC users “can delegate everyday projects … without sending those workloads to the cloud.” For privacy-conscious users who will not hand a cloud agent access to their family, health, and financial data, a capable local agent is the product they have been waiting for, and it is a product that sells Nvidia chips.
What it means if you run agents at work
If your company is already experimenting with always-on agents for document processing, triage, or research, Nemotron 3.5 gives you a new option between two extremes. On one side are closed frontier APIs, billed per token, that you cannot post-train and cannot run on your own servers. On the other are smaller open models that cannot sustain long multi-step work. Super VL is pitched at the gap: big enough for serious agent work, open enough to own.
The honest caveat is that Nvidia’s speed claims are self-reported and task-specific, and independent testing on Artificial Analysis and elsewhere will be the real verdict. But the direction is unmistakable. The agent era is moving from chatbots that answer questions to software that does work, and the fight is shifting from who has the smartest model to who controls the infrastructure it runs on. Nvidia, characteristically, wants to sell you the infrastructure either way.
Related: ChatGPT Can Now Read Your Chrome Tabs: Tab Comparison, Work Mode, and the Permission Checks to Make First, DeepSeek V4.1-Flash-1017 Beats Claude Sonnet 5.6 on 10 of 11 Coding Tests While Using 20 Percent Fewer Tokens, OpenAI Halved ChatGPT Pro: the $200 Plan Drops to 10x Usage on October 30, Plus a $500 Tier
