Two models, both built to be cheap to run
Nvidia has released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters of which 12 billion are active per token, and Nemotron 3.5 Bolt 74B-A9B 24GB, an edge model with 74 billion total and 9 billion active that uses native two-bit weight quantisation to fit inside 24GB of memory. Nvidia calls Super VL the fastest open agentic model in its class and says it is around five times faster than the next-best model in that class.
Super VL handles text, images, documents and video. The Decoder reports that it was distilled from Nvidia’s larger Nemotron 3 Ultra and uses the company’s four-bit NVFP4 format, running fully in four-bit precision on Blackwell GPUs and storing weights at four bits while computing at sixteen on older cards.
The numbers Nvidia gives, and the one it does not
Nvidia says Super VL is comparable to its largest reasoning model at about a fifth of the parameter count, and that Bolt 74B ranks first among open models in its class on PinchBench and second on BrowseComp. It gives no numeric scores for either. On the independent Artificial Analysis Intelligence Index, The Decoder reports Super VL at 26 and Bolt at 21 — above Nemotron 3 Ultra, and a long way below the closed frontier, with Claude Opus 5.5 at 58 on the same chart.
The serving claims are more concrete. Nvidia says that with its Dynamo serving stack, Super VL handles up to eight times more concurrent users with 85% lower latency. Training took roughly 26 trillion tokens of pretraining and continued pretraining, followed by about 3.8 million post-training samples and reinforcement learning across dozens of environments.
Where it runs, and who is already using it
Super VL is on Hugging Face, OpenRouter, the ModelScope community and as an Nvidia NIM microservice. Bolt 74B is on Hugging Face and runs through llama.cpp, Ollama, LM Studio, Unsloth and vLLM on GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations. A smaller Bolt E4B 4GB aimed at voice agents is due on 23 October, and Nvidia says the models are built for the RTX Spark laptops arriving on 16 October.
Nvidia names a long list of companies post-training or deploying the models, including Scale AI, AI21, Mercor, Ramp, Lila Sciences, Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys. The Decoder reports that Super VL was developed with a Nemotron Coalition that includes Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs.
What to watch
Nvidia publishes weights, datasets and training recipes, so the claims above are testable in a way vendor benchmark charts are not. Jensen Huang’s GTC Berlin keynote on 21 October is the next place the Nemotron line is likely to move.
