NVIDIA said on Oct. 15 that Nemotron 3.5 Super VL, an open model with 120 billion parameters, is available now. So is Bolt 74B-A9B, a model for local use. A third, Bolt E4B, is due Oct. 23. NVIDIA said Super VL reads web pages, PDFs, charts and tables, and that it is sharing the weights plus the data and methods behind it. Bolt 74B-A9B runs with Ollama, LM Studio and vLLM.
NVIDIA said Bolt 74B-A9B fits in 24GB of memory using 2-bit quantization, which stores a model's numbers in very few bits. The post does not say what accuracy that costs. NVIDIA also said Super VL is five times faster than the next best open model in its class. It said Bolt ranks first on PinchBench and second on BrowseComp among open models in its class. The post gives no outside tests, prices or licence terms. It names firms adapting Super VL, including Scale AI and CrowdStrike, but not which run it in production.
NVIDIA's pitch is that open weights let firms run agents on their own hardware and keep data on site. Buyers can ask for licence terms and production customers, test the models on their own tasks, and price the hardware needed.