The key facts

  • Nemotron 3.5 Super VL — 120 billion parameters for enterprise agents; Nvidia claims accuracy comparable to Nemotron 3 Ultra with fewer parameters
  • Two Bolt models for local execution: 74B-A9B 24GB (74 billion parameters) and the compact E4B for conversational agents
  • Nvidia claims Super VL is faster than competitors Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra) in post-training for agentic tasks; named users include Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce

A new lineup of open models

On Thursday, October 19, Nvidia introduced new open-weight models aimed at always-on agents and edge computing. The flagship Nemotron 3.5 Super VL contains 120 billion parameters and is designed for corporate users. According to Nvidia, with fewer parameters than Nemotron 3 Ultra, it delivers comparable accuracy in instruction-following and tool-use tasks.

Models for local AI PCs

At the same time, two lightweight Bolt models were released, intended to run locally on new AI PCs with the Nvidia RTX Spark chip and on DGX Spark workstations. Such computers, aimed at consumers, programmers, and office workers, are already reaching the market through partners including Microsoft and Dell. The Nemotron 3.5 Bolt 74B-A9B 24GB model (74 billion parameters) is built for local agentic tasks: multi-step reasoning, tool calling, and web browsing. The most compact one, Bolt E4B, targets conversational agents. Super VL and Bolt 74B-A9B are available immediately; Bolt E4B will appear later this month.

Performance claims and customers

Nvidia claims that Super VL, after post-training for specific agentic tasks — document analytics, financial accounting, scientific experiments, medical insurance claims processing, and attack resilience — performs them faster than competing models from Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra). In support of these claims, the company lists major users of Super VL: Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.

Nvidia's strategy

Nvidia's vice president of open models and software, Kari Briski, wrote in a blog post that organizations wanting to run agents on Nvidia infrastructure get 'an open, optimized model for their tier without choosing between capability and deployability.' According to her, unlike closed frontier models, companies can post-train Super VL on their own data, deploy it on their own infrastructure, and own the result. This is part of Nvidia's bet on the trend of running open models in companies' own data centers: fast-growing proprietary AI labs are increasingly turning to other chip makers, such as AWS Trainium and Cerebras, and are also developing their own silicon. Nemotron is already an entire family of Nvidia models: from multimodal AI to reasoning, code, math, and enterprise automation.

A bet on local consumer assistants

With local PC models, Nvidia is also promoting the emerging trend of consumer AI assistants. The company expects that users concerned about privacy and unwilling to hand assistants data about family, health, and finances will choose machines that run models and agents locally: according to Briski, users 'can delegate everyday projects without sending those tasks to the cloud.' In doing so, Nvidia hopes to create a new market for itself in PCs built on its own chips.