NVIDIA publishes NVFP4 Nemotron 3.5 Super VL 120B model
Original title: nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 — new model trending #30 on Hugging Face
The open-weight vision-language release has 12B active parameters, up to 1M-token context, and OpenMDW-1.1 terms.
NVIDIA released NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4, an NVFP4 quantized version of Nemotron 3.5 Super VL distilled from the BF16 model. It targets reasoning, coding, tool use, and document and video understanding, with 120B total and 12B active parameters and context up to 1M tokens. The card lists deployment on one DGX Spark, one H200, two H100s, or one GB200, plus a vLLM quick start. NVIDIA reports NVFP4 scores close to BF16 on GPQA-Diamond, text-only HLE, and MMLU-ProX.
Unconfirmed: Benchmark figures are NVIDIA's own and the listing is truncated.
- Architecture mixes Mamba-2, MoE, attention, and a C-RADIO vision encoder.
- Inputs are text, images, and video; outputs are text.
- Native Multi-Token Prediction supports speculative decoding.
- Recommended sampling differs for text and vision.
- The model card says it is ready for commercial or non-commercial use.
- Vendors
- NVIDIA
- Organizations
- NVIDIA Corporation
From the original article
11 min read at huggingface.coimage-text-to-text · 9,322 downloads · 165 likes · license other
By nvidia · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.