Nvidia launches Nemotron 3.5 Super VL and Bolt open models

Nvidia launched two new families of open-weight models on Thursday, targeting both enterprise data centers and the emerging market of local AI PCs. The flagship release, Nemotron 3.5 Super VL, is a 120-billion-parameter model designed for always-on agentic workflows, while two smaller "Bolt" models are optimized to run locally on consumer hardware like the new RTX Spark and DGX Spark workstations. TechCrunch reports that Nvidia is positioning these releases as a direct challenge to proprietary models from Anthropic and OpenAI, offering organizations the ability to post-train and deploy agents on their own infrastructure without sending data to the cloud.

The details

The Nemotron 3.5 Super VL model uses fewer parameters than Nvidia’s existing Nemotron 3 Ultra but claims similar accuracy in instruction-following and tool-use tasks. Kari Briski, Nvidia’s vice president of Open Models and Software, stated in the announcement that the model is optimized for specific agentic tasks, including document intelligence, financial accounting, scientific experimentation, and medical insurance claims. Nvidia claims Super VL completes these tasks faster than competing models, specifically naming Anthropic’s Opus 5 and OpenAI’s GPT 6.1 Sol, Luna, and Astra, when post-trained on these verticals.

On the edge computing side, Nvidia introduced two lightweight Bolt models. The Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model designed for local agentic tasks, balancing speed with reasoning to handle multistep reasoning, tool calling, and web browsing. The smaller Bolt E4B is aimed at conversational agents. Both models are intended to run on AI PCs powered by the Nvidia RTX Spark chip and workstations running the DGX Spark, which are being released through partners like Microsoft and Dell.

Nvidia lists several heavy-hitter companies as early adopters of Super VL, including Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce. Briski emphasized that unlike closed frontier models, organizations can post-train Super VL on their own data and own what they build. The Super VL and Bolt 74B-A9B models are available now, with the Bolt E4B scheduled to arrive later this month.

Context

This launch continues Nvidia’s strategy of releasing open-weight models that compete with the proprietary frontier labs that are both its customers and its investors. The Nemotron family has previously covered multimodal AI, reasoning, coding, and math. However, the competitive landscape is shifting as AI labs look beyond Nvidia to power their needs, utilizing chips from AWS’s Trainium, Cerebras, and building their own home-grown silicon. Nvidia is leaning into the trend of enterprise companies wanting to run their own open models in their own data centers to mitigate dependency on external API providers. Simultaneously, the company is pushing the nascent trend of consumer AI assistants, aiming to create a new market for PCs that run models and agents locally, addressing privacy concerns among consumers who may not want to invite cloud-based agents into their personal lives.

What it means for developers

For developers and enterprise architects, the primary value proposition is the ability to fine-tune a high-performance agent model on proprietary data without licensing constraints. The Super VL model is specifically optimized for tool-use and instruction-following, making it a viable alternative for building internal copilots in finance, healthcare, or security. For local development, the Bolt 74B-A9B offers a path to running complex agentic workflows on a single workstation, provided the hardware supports the 24GB memory requirement. Developers should note that while Nvidia claims parity with frontier models on specific tasks, independent benchmarks are still needed to verify performance in general-purpose coding and reasoning scenarios. The availability of these models on local hardware also opens up opportunities for offline-first applications, which is critical for industries with strict data residency requirements.

What to watch

• Independent Benchmarks: Look for third-party evaluations of Nemotron 3.5 Super VL against GPT 6.1 and Opus 5 to verify Nvidia’s speed and accuracy claims.
• Bolt E4B Release: The arrival of the smaller conversational model later this month will test the viability of ultra-lightweight local agents.
• Hardware Adoption: Monitor the market uptake of RTX Spark and DGX Spark workstations to gauge the success of Nvidia’s local AI PC strategy.
• Competitor Response: Watch for reactions from Anthropic and OpenAI, who may adjust their pricing or release new open-weight models in response to Nvidia’s enterprise push.