Nvidia releases Nemotron 3.5 Super VL, an open 120-billion-parameter vision model, with a new Indic-language dataset
The model reads text, images, documents and video and is built to be post-trained for specific business tasks. Nvidia also released about 2.8 trillion tokens of Indic text across 22 languages.
Illustration · not a photoHow this story was checked
- ResearchedWith AI assistance · 3 sources
- AI source checkEach sentence matched to its source
- Editor approvalMayank Sahu · 16 Oct, 13:37
- Published16 Oct, 13:37
Verify this story
- 3 sources, 2 official · 2 with an archived copy · see sources
- Every fact has a label and a source number · fact ledger
- What we searched, read and could not verify · research note
- Machine-readable: facts as JSON · text as Markdown
How this was made: researched, drafted and source-checked by AI (Claude, made by Anthropic), then published under the editor’s standing approval. A person did not read it line by line before it went live. AI policy
In short
Nvidia released Nemotron 3.5 Super VL on 15 October 2026, an open-weight mixture-of-experts model with 120 billion total and 12 billion active parameters that accepts text, images, documents and video. Nvidia also released an Indic dataset of about 2.8 trillion tokens across 22 languages.
What Nvidia released
Nvidia published Nemotron 3.5 Super VL on 15 October 2026.1 It is a mixture-of-experts model with 120 billion total parameters and 12 billion active for each token, using a hybrid design that includes Mamba layers.1 It accepts text, images, documents and video.1
Nvidia released base, post-trained and compressed NVFP4 checkpoints, along with training recipes and datasets.13 VentureBeat reports the licence, OpenMDW, allows commercial use, fine-tuning and redistribution.2
Why it matters for India
Nvidia says it added 10 billion Indic-language tokens in continued pretraining and released an Indic dataset of about 2.8 trillion tokens across 22 languages.1 That gives developers building for Indian languages a large, openly available training set.
Where to get it
The weights are on Hugging Face, with access also through ModelScope, OpenRouter and Nvidia NIM.13 VentureBeat reports a free, rate-limited Nvidia-hosted API for development, and that production NIM use needs an Nvidia AI Enterprise licence, quoted at $4,500 per GPU per year.2
What Nvidia claims
- 80.8% on the CharXiv chart-reasoning test, or 84% with a Python tool.1
- More than 600 output tokens a second on agent tasks through a Baseten endpoint.1
- After task-specific post-training, accuracy on a medical insurance claims test rose from about 28% to 72%, against about 38% for Claude Opus 5, VentureBeat reports from Nvidia’s charts.2
VentureBeat notes the comparisons use customised models on narrow tasks, and Nvidia has not published task definitions or sample sizes.2
Also announced
Nvidia also introduced Nemotron 3.5 Bolt 74B-A9B for local coding and research, and VentureBeat reports a smaller Bolt E4B model for voice and device agents is due on 23 October.2
Fact ledger
How labels workTracked from this story
All trackersRelease the Nemotron 3.5 Bolt E4B model for voice and device agents
Pending Follow itQuestions readers ask
Is Nemotron 3.5 Super VL free?
The weights are open. VentureBeat reports the OpenMDW licence allows commercial use; Nvidia’s supported production service needs a paid licence.
Does it support Indian languages?
Nvidia says it added Indic-language training data and released a 22-language Indic dataset of about 2.8 trillion tokens.
Is it better than Claude or GPT?
Nvidia’s comparisons cover custom post-trained versions on narrow tasks. They do not show the base model is better in general.