Nvidia launches open-weight Nemotron models for enterprise agents and AI PCs

Nvidia on Thursday released a 120-billion-parameter open-weight model, Nemotron 3.5 Super VL, built for always-on enterprise agents, alongside two smaller Bolt models designed to run locally on consumer AI PCs. The release marks Nvidia's most direct push yet into two markets at once: companies that want to run their own models in their own data centers, and consumers who want agents that never leave their machine.

The stakes are straightforward. Nvidia's fastest-growing customers — the proprietary frontier labs — are diversifying their chip suppliers, turning to AWS's Trainium, Cerebras, and home-grown silicon. Open-weight enterprise models give Nvidia a second revenue lane that does not depend on frontier-lab training runs. The company has an existing Nemotron family covering multimodal AI, reasoning, coding, math, and enterprise automation; Thursday's launch narrows that portfolio toward agentic workloads, the current center of enterprise spending.

What did Nvidia actually release?

Three models, spanning roughly two orders of magnitude in size:

• Nemotron 3.5 Super VL — a 120-billion-parameter model for enterprise agents. It uses fewer parameters than the Nemotron 3 Ultra model, but Nvidia says it delivers similar accuracy on instruction-following and tool-use tasks (the core abilities an agent needs to execute multi-step work reliably).
• Nemotron 3.5 Bolt 74B-A9B 24GB — a 74-billion-parameter model for local agentic tasks, built to balance speed with reasoning across multistep reasoning, tool calling, and web browsing.
• Bolt E4B — the smallest of the three, aimed at conversational agents.

Super VL and Bolt 74B-A9B are available now. Bolt E4B arrives later this month, Nvidia says.

Who is the enterprise pitch for?

Kari Briski, Nvidia's vice president of Open Models and Software, framed the release as a tier-by-tier offer. "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," she wrote in the announcement blog post.

The emphasis on ownership is deliberate. "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," Briski wrote. Post-training (fine-tuning a released model on domain-specific data) is the operational argument for open weights: a bank or insurer can specialize the model without shipping sensitive data to a third-party API.

That framing targets a real shift. Enterprise companies increasingly want self-hosted open models in their own data centers, and Nvidia is leaning into that demand as frontier labs — companies that are simultaneously Nvidia's customers and the beneficiaries of its investments — spread their compute across alternatives.

What does Nvidia claim about performance — and what backs it up?

Nvidia claims Super VL completes tasks faster than competing models from Anthropic and OpenAI — specifically Opus 5, GPT 6.1 Sol, Luna, and Astra — when post-trained on specific agentic tasks. The listed task domains include document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience.

Vendor benchmark claims at launch are standard practice for major model releases, and this one carries the usual caveat: the results depend on Nvidia's own post-training setup, and the comparisons reflect Nvidia-chosen tasks. What gives the claim more weight than a typical launch slide is the customer list. Nvidia says Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce are using Super VL. Named deployments are a stronger signal than benchmark tables, though Nvidia did not disclose how those companies are using the model or at what scale.

The takeaway: the performance claims are unverified by third parties, but six recognizable names have publicly attached themselves to the model.

Why put agents on a PC?

The two Bolt models serve a different bet. They run locally on the new AI PCs built around the Nvidia RTX Spark chip and on workstations running the DGX Spark. Microsoft and Dell are releasing these machines now, aimed at consumers, programmers, and enterprise knowledge workers.

The local pitch is privacy. A new AI assistant seems to appear daily from startups and big tech alike, but agents that manage families, health, and finances demand deep access to personal information — access many privacy-wary users may resist granting to a cloud service. Nvidia's answer is to move the workload onto the machine. PC users "can delegate everyday projects … without sending those workloads to the cloud," Briski wrote.

The commercial logic is clear: if local agents become a reason to buy a PC, Nvidia sells more chips into a hardware category where it has historically been a GPU supplier rather than a platform owner. The caveat mirrors the enterprise one — the nascent consumer-agent market has not yet demonstrated sustained demand, and local models face hard memory and thermal limits that cloud models do not.

What is the strategic picture?

Nvidia is hedging against customer concentration. The frontier labs that drove its data-center growth are now sourcing chips from AWS (Trainium), Cerebras, and their own internal designs. Open-weight models sold to enterprises — and, now, agents shipped on consumer hardware — diversify Nvidia's AI footprint beyond that increasingly contested relationship.

There is tension in the strategy worth noting. Nvidia's open models compete directly with the proprietary labs it funds and supplies. The company appears to have concluded that the enterprise self-hosting market is large enough to justify that friction.

Super VL and Bolt 74B-A9B are downloadable now; Bolt E4B follows later this month. Whether the enterprise deployments scale beyond the named early adopters — and whether consumers buy AI PCs to run local agents at all — will determine whether this launch is a platform shift or a portfolio expansion.