NVIDIA releases Nemotron 3.5 Super VL 120B with 1M-token context and speculative decoding
NVIDIA launched the Nemotron 3.5 Super VL 120B-A12B-BF16 model, a vision-language model supporting text, images, and video inputs. It combines Mamba-2, Mixture-of-Experts (MoE), and attention layers with a C-RADIO vision encoder, enabling reasoning, coding, tool use, and long-context understanding. The model has 120B total parameters and 12B active parameters, with a 1M-token context window and…
Key points
- Nemotron 3.5 Super VL 120B supports **1M-token context** with **text, image, and video inputs** via MTP speculative decoding
- Model combines **Mamba-2, MoE, and attention layers** with a **C-RADIO vision encoder** for multimodal reasoning
- Optimized for **NVIDIA Hopper/Blackwell GPUs**, released under **OpenMDW License 1.1** on Hugging Face
The model is optimized for NVIDIA GPUs (Hopper/Blackwell) and released under the OpenMDW License 1.1. NVIDIA provides deployment guides for vLLM and SGLang, with examples for text, image, video, and tool-calling tasks. Benchmarks cover agentic capabilities, reasoning, multimodal tasks, and multilingual performance, though full validation requires use-case-specific testing. The model is available on Hugging Face as of October 15, 2026, targeting developers for customization, domain adaptation, and inference.
The story so far
2 episodes →- NVIDIA releases Nemotron 3.5 Super VL 120B with 1M-token context and speculative decodingthis story
nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B model family released (base, NVFP4, BF16, and LiLiCorr speculative decoding)
huggingface.co · 12 October 2026
Loading the full article…
This text was published by huggingface.co. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
2sources- Reddit discussionreddit.com
- Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt ModelsPrimary source · NVIDIA Blog ·
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- AtomicChat releases quantized Ling 3.1 flash models with 86–88% token accuracy · 1 src
- OpenAI launches GPT-6.1 Sol with Astra-level coding at one-fifth the cost · 2 src
- Add Nemotron Super 3.5 VL and Nemotron Bolt 74B A9B support by frozenblade1224 · Pull Request #30520 · ggml-org/llama.cpp · 1 src
- Apertus 1.5 adds 4T tokens to 8B model, 2T to 70B, with multimodal and long-context upgrades · 1 src
- Clash of AIs platform shows ChatGPT leads Claude in market tests · 1 src
Comments
via GitHub Discussions