Nemotron 3.5 Super VL

Not yet independently verified. Every detail here comes from NVIDIA's own model card, which dates the release to 15 October 2026; no independent report was found. NVIDIA's benchmark tables are its own evaluation and no score is stored. No hosted API price is stated. We will update this when it can be confirmed, and remove this note.

Nemotron 3.5 Super VL is a vision-language model from NVIDIA, published on Hugging Face on 15 October 2026 with full-precision BF16 weights. It reads text, images and video and returns text, and NVIDIA positions it for reasoning, coding, tool use, document understanding and video understanding. The model has 120 billion parameters in total, of which 12 billion are active for each token.

What it does

The architecture interleaves Mamba-2 layers, mixture-of-experts layers and attention layers, with a C-RADIO vision encoder and a connector that turns images and video frames into input for the language backbone; a shared multi-token prediction head speeds up generation. The context window reaches one million tokens, shared between the prompt, visual inputs, conversation history, tool results and the answer, although NVIDIA notes that individual deployment recipes may use less. Documents can be supplied as page images with accompanying text. NVIDIA's model card describes training stages that include 100 billion tokens of Indic continued pre-training, an extension to a one-million-token context and 67 billion tokens of software-engineering training, followed by supervised fine-tuning and reinforcement learning. The card lists English, Spanish, French, German, Italian and Japanese among supported languages.

Who it is for

NVIDIA says the BF16 release is meant for developers who want to customise the model, adapt it to their own data, build quantised variants or run full-precision inference and evaluation. The licence allows commercial and non-commercial use, and NVIDIA lists deployment as global. The model card includes NVIDIA's own benchmark tables across agentic, reasoning, multimodal, long-context and multilingual tests; those are the vendor's figures and no score is stored here.

Key facts

Maker: NVIDIA Corporation Released: 15 October 2026, on Hugging Face Size: 120B total parameters, 12B active Context: up to 1M tokens Inputs: text, images, video; output text Licence: OpenMDW-1.1, commercial and non-commercial use