Nvidia releases open Nemotron 3.5 Super VL, a 120B multimodal model, and 2-bit Bolt for local agents

Super VL reads PDFs, charts, tables, images and webpages while following instructions and calling tools. Nvidia pitches it for post-training on a company's own data, with deployment where security and data-residency rules require. Bolt 74B-A9B activates 9 billion of 74 billion parameters and, using native 2-bit compression, fits in 24GB to run on RTX Spark and DGX Spark systems. The 4GB Bolt E4B is due October 23. Nvidia's charts show post-training gains of 9 to 44% across five tasks. Medical claims rose from roughly 28% to 72%, against about 38% for Claude Opus 5, and security resilience from 35% to 76%, against about 66% for GPT-6.1 Sol. VentureBeat notes these are customized models on narrow tasks, and Nvidia gave no task definitions or sample sizes. Super VL reaches about 625 output tokens per second on four B200s, but its Artificial Analysis score of 26 trails GLM-5.3-Flash at 42 and Qwen3.8-Flash-Next at 40. For builders, this is an open, tunable document-agent base with a local option. Test it on your own documents before trusting the vendor charts.