THE CRUNCH
NVIDIA has released Nemotron 3.5 Super VL, an open 120-billion-parameter vision-language model built for specialised agents that run continuously in the background. It handles the mix of work agents actually meet: instruction following, tool use, multistep workflows, and reading web pages, PDFs, tables and charts. Two compact Bolt models ship alongside it: Bolt 74B-A9B 24GB for local agentic tasks on RTX and DGX Spark hardware, and Bolt E4B 4GB for fast conversational and voice agents on GeForce RTX, Jetson and Spark. Super VL and Bolt 74B-A9B are available now; Bolt E4B is due on 23 October.
Super VL is a Mixture-of-Experts model, meaning only 12 billion of its 120 billion parameters activate per token, and NVIDIA distilled it from Nemotron 3 Ultra, its largest reasoning model. The company claims it matches Ultra's accuracy at roughly one-fifth the size, and cites the Artificial Analysis leaderboard for up to five times the throughput of comparable open models on agent tasks. In a test the company points to from EXO Labs on the DGX Spark, no other model was both more accurate and faster.
The efficiency push extends to serving. Super VL caches the computations for instructions and conversation history that agents resend with every request, and NVIDIA's Dynamo software routes repeat sessions to the GPU holding that cache. On 16 B200 GPUs, the company says this serves eight times as many concurrent users and cuts time to first token by 85 percent.
Because the weights, datasets and training recipes are open, organisations can post-train the model on their own data and run it on their own infrastructure. NVIDIA lists dozens already doing so, including Scale AI for Skydance's vendor credit finance agent and EY's M&A due diligence agent, Atlassian for agentic search across Jira and Confluence, and CrowdStrike for enterprise security work.