Nvidia releases open Nemotron 3.5 Super VL (120B MoE, 12B active) multimodal agent model, plus two smaller Bolt models for local hardware

Nvidia has published open weights, datasets and training recipes for Nemotron 3.5 Super VL, a 120B-parameter Mixture-of-Experts multimodal model that uses 12B parameters per token. It is distilled from the larger Nemotron 3 Ultra and aimed at long-running enterprise agents and document and vision tasks. Nvidia also released Nemotron 3.5 Bolt 74B-A9B, a 2-bit version that fits in 24 GB of memory, with a 4 GB Bolt E4B voice model due later this month. The models are compressed in NVFP4, tuned for Nvidia hardware and Dynamo, and work with llama.cpp, Ollama, LM Studio and vLLM.

Developers and enterprises can download, fine-tune and self-host these models now, which helps teams that want agents without closed-provider terms or data leaving their own systems. The datasets and training recipes are the more unusual part, because they let others reproduce or adapt the work. The practical gain is limited by performance: the coverage admits the models trail leading closed models by a wide margin, and the local Bolt model still needs costly hardware. It is a useful option in a crowded open-model field, not a shift in who leads.

The headline claims, matching the much larger Nemotron 3 Ultra on agent tasks and up to 5x throughput, come mostly from Nvidia's own charts, and the throughput figure depends on Nvidia's hardware and software stack. Only one independent test (EXO Labs) is mentioned. The "open" framing also serves Nvidia's commercial interest in selling chips and systems, and the coalition names add prestige without showing adoption.

1. 01Independent benchmark results from Artificial Analysis and others on agent and document tasks, compared with Nemotron 3 Ultra and leading open rivals, especially on hardware that is not Nvidia's.
2. 02Whether the Bolt E4B 4GB voice model ships on schedule later this month, and what licence terms apply to all three models.
3. 03Named production deployments (for example Scale AI, Ramp or EY) and download and fine-tune counts over the next one to three months, which would show real uptake beyond launch publicity.