NVIDIA Nemotron 3.5 Super VL 120B

• 🏢 NVIDIA open-weight vision-language model released October 2026.
• 🧠 120B total parameters, 12B active per token (MoE).
• 🆕 Hybrid design: Mamba-2, attention, MoE, C-RADIO vision encoder.
• 👁️ Handles text, images, documents, GUIs and video in one checkpoint.
• 📏 Model card cites up to 1M-token context; served here at 262K.
• ⚡ Multi-Token Prediction enables native speculative decoding.
• 🔧 Built for tool use, coding and long-running agentic workflows.
• 📚 Weights published in BF16, NVFP4 and base variants.

Nvidia

Nvidia Corporation is an American technology company founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, headquartered in Santa Clara, California. Long recognized as the dominant force in graphics processing units, Nvidia has expanded into a…

NVIDIA Nemotron 3.5 Super VL is the multimodal "Super" tier of NVIDIA's open Nemotron line, aimed at developers building specialized agents that read documents, screens, charts and video while also reasoning, calling tools and writing code. Rather than pairing a text model with a separate vision model, it folds all of that into a single open checkpoint.

Architecturally it is a Nemotron Hybrid Mixture-of-Experts: interleaved Mamba-2, attention and MoE layers with 512 routed experts per MoE layer, plus a C-RADIO vision encoder and a vision-language connector. Of its 120B total parameters, 12B are active per token, and Multi-Token Prediction provides native speculative decoding. The Mamba-2 and sparse-expert mix is what lets it process very long inputs; NVIDIA's model card states support for context lengths up to 1M tokens, while this catalog entry exposes a 262K window in an FP4 quantization.

Within the family, the model sits alongside the text-only Nemotron 3 generation, including the flagship NVIDIA Nemotron 3 Ultra and the compact Nemotron 3 Nano 30B. The main generational additions in this 3.5 Super VL release are native visual, document and video understanding fused into the language backbone, together with post-training oriented toward agentic tool use and long-running workflows.

NVIDIA publishes the weights in several forms on Hugging Face: a BF16 instruct checkpoint, an NVFP4 quantized checkpoint for efficient serving, and a BF16 base checkpoint for further training. For retrieval pipelines it can be paired with Nemotron Embed VL 1B v2.