Nvidia released Nemotron 3.5 Super VL on 15 October, a multimodal mixture-of-experts model with 120 billion total parameters and 12 billion active, built on a hybrid Mamba-Transformer backbone with a 1 million-token context window. Weights ship under the OpenMDW licence, with BF16 and NVFP4 builds on Hugging Face.
Two smaller variants landed alongside it: Bolt 74B-A9B and an E4B edge build. Nvidia claims up to 5x the throughput of comparable open models, which is the hybrid backbone doing the work — Mamba layers scale linearly with sequence length where attention scales quadratically, so the advantage grows with context used rather than being a flat multiplier.
The positioning is agents, not chat. A 12B active footprint with vision and a million tokens of context is aimed at long-running agent loops that read screens and documents, where per-token cost compounds over hundreds of steps. The 5x figure is Nvidia's own; independent throughput reproductions were not available at publication.
Sources:
Hugging Face ·
MarkTechPost ·
Baseten