NVIDIA releases Nemotron 3.5 Super VL and Bolt
On 15 October 2026 NVIDIA released Nemotron 3.5 Super VL, a 120-billion-parameter open vision-language model for agents, and two edge models: Nemotron 3.5 Bolt 74B-A9B 24GB, available the same day on Hugging Face, and Nemotron 3.5 Bolt E4B 4GB, due on 23 October. NVIDIA says the models ship with open weights, open datasets and training recipes, and that Super VL is up to five times faster than the next best model in its class on the Artificial Analysis leaderboard.
Why it matters
NVIDIA, a chip maker, now fields a full line of open models aimed at agents, from a data-centre model to one sized for a 24 GB graphics card, and releases datasets and training environments with it. The page describes the weights as open without naming a licence, so what a user may do with them is not stated there. The speed and accuracy claims are NVIDIA's reading of a third-party leaderboard and of its own benchmarks.
What the page says. Blog post of 15 October 2026 by Kari Briski. Super VL (120 billion parameters), built with the Nemotron Coalition, is for multimodal agent tasks (instruction following, tool use, multistep business workflows, visual understanding; OCR, tables, charts and PDF pages) and is said to match NVIDIA's largest reasoning model "at roughly one-fifth the parameter count"; on the Artificial Analysis Intelligence Index leaderboard it "achieves up to 5x faster throughput compared to other models in its class". It was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained on roughly 3.8 million samples and with reinforcement learning across dozens of environments. It is available on Hugging Face, OpenRouter, the ModelScope community and as an NVIDIA NIM. Bolt 74B-A9B 24GB uses native 2-bit quantisation to fit in 24 GB and run locally on RTX and DGX Spark; NVIDIA calls it the #1 open model in its class on PinchBench and #2 on BrowseComp; it is available on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth, Canonical and vLLM. Bolt E4B 4GB is for conversational agents on GeForce RTX, Jetson and DGX Spark and is "coming Oct. 23". The page says the family also includes Lightning (released in August) and Ultra. It lists organisations that customise the model (Scale AI for Skydance and EY, AfterQuery for Hebbia, Mercor and others) and software companies that integrate it (Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce, Synopsys); in a heading it writes "Custom Super Intelligence" where it means AI tailored to a customer. NVIDIA is also releasing open datasets on Hugging Face (agentic tool use, finance, instruction following, long context, fairness, vision-language understanding) and reinforcement-learning environments in NeMo Gym. What the record does not claim. No licence is named (the page says only "open weights, open datasets and training recipes"); the rankings and speed multiples are NVIDIA's own words, and no independent test was read; the benchmark charts were not read; whether Bolt E4B was released on 23 October was not checked (the page says it is coming).