NVIDIA announced on Oct 15, 2026 that three new open models are now available. The flagship Nemotron 3.5 Super VL has 120 billion parameters and is marketed as the fastest and most accurate open model for always‑on agents. Two edge‑focused variants, Nemotron 3.5 Bolt 74B‑A9B (24 GB) and Bolt E4B (4 GB), are also released, with the smaller model slated for launch on Oct 23, 2026. All three can be downloaded from Hugging Face, Open Router and ModelScope, and run on NVIDIA’s DGX and RTX hardware.
What the new models claim to do
Super VL is said to run five times faster than the next open model in its class and to match the accuracy of NVIDIA’s largest reasoning model while using only one‑fifth as many parameters. The 74B‑A9B Bolt model uses 2‑bit quantization to fit in 24 GB of GPU memory, letting it run locally on RTX‑based laptops and workstations. The 4 GB Bolt model targets ultra‑light devices such as Jetson and desktop RTX cards for fast conversational agents.
Who is already using the models
NVIDIA lists dozens of companies that have fine‑tuned Super VL on private data, including Scale AI for finance agents, AfterQuery for credit‑agreement analysis, and TrainLoop for dermatology prescription drafting. Platform partners such as Atlassian, CrowdStrike, Salesforce and Perplexity are integrating the model into their own AI‑agent products.
Why it matters
For businesses that need on‑premise AI agents, the open weights let them keep data private while still getting high‑speed performance. For developers, the smaller Bolt models mean they can run sophisticated agents on a laptop or edge device without relying on cloud services.
