NVIDIA Releases Nemotron 3.5 Super VL and Bolt Edge Models

NVIDIA has released Nemotron 3.5 Super VL, an open 120-billion-parameter model for specialised always-on agents, together with two edge models: Nemotron 3.5 Bolt 74B-A9B 24GB, for agentic tasks run locally, and Nemotron 3.5 Bolt E4B 4GB, for conversational agents.

Super VL and Bolt 74B-A9B are available now; Bolt E4B is due on 23 October. They join Nemotron Lightning, released in August for high-throughput work, and Nemotron Ultra for complex reasoning.

Super VL

NVIDIA says Super VL matches the accuracy of its largest reasoning model on multimodal agentic tasks at roughly one-fifth of the parameter count, and that it is 5x faster than the next-fastest open agentic model in its class. Served on NVIDIA Dynamo, the company puts the gain at up to 8x more users with 85% lower latency.

The model was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained on roughly 3.8 million samples with reinforcement learning across dozens of environments. Weights, datasets and training recipes are open. It is available on Hugging Face, OpenRouter and ModelScope, and as an NVIDIA NIM microservice that NVIDIA refreshes every two weeks.

Bolt

Bolt 74B-A9B uses native 2-bit quantisation to fit in 24GB of memory. It runs with llama.cpp, Ollama, LM Studio, Unsloth and vLLM, and is built for the NVIDIA RTX Spark laptops that arrived on 16 October; it also runs on GeForce RTX PCs, DGX Spark and DGX Station. NVIDIA lists it as the top open model in its class on PinchBench and second on BrowseComp.