NVIDIA Releases Nemotron 3.5 Super VL, An Open-Weight Model For Text, Images, And Video
The Skydance case is the part worth watching: a tuned open-weight model holding accuracy while cutting processing costs to about one-fortieth reframes what the open-weight tier can be sold for against closed API pricing.
Reporting from 1 source: GIGAZINE.
NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, an open-weight model that reads text, images, and video in one system and operates external tools. It carries 120 billion parameters and, through mixture-of-experts routing, uses 12 billion per token. NVIDIA claims up to five times the throughput of open models at the same scale. Scale AI tuned a version for Skydance that cut processing costs to about one-fortieth.
The parameter count is the smaller number. Of the 120 billion parameters in Nemotron 3.5 Super VL, mixture-of-experts routing keeps 12 billion active per token, and NVIDIA reports throughput above 600 tokens per second. On 10,000 software development tasks across four B200 GPUs, the model finished in about 10 hours at roughly 51% accuracy, up to four times faster than other models at the same scale by NVIDIA's account.
Graph reading landed at 80.8% on CharXiv, rising to 84% when Python is used to analyze the charts, against 63% for NVIDIA's Nemotron 3 Nano Omni.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.