AgentPlays
NVIDIA releases Nemotron 3.5 Super VL, an open 120B multimodal model for long-running agents
On Oct 15, NVIDIA released Nemotron 3.5 Super VL, an open vision-language model with 120B total parameters and 12B active per token. It reads text, images, documents and video, writes text, and supports up to a 1M-token context. It's a hybrid Mamba-2 and Transformer mixture-of-experts model with a C-RADIO vision encoder and multi-token prediction for faster decoding. BF16 and NVFP4 weights are on Hugging Face under OpenMDW-1.1, a permissive license that allows commercial use with no region limits. NVIDIA also released open training datasets and recipes, and says teams can post-train it with NeMo using fine-tuning, LoRA and reinforcement learning. vLLM shipped day-zero support the same day.
How you’d use it: If you're building an agent that reads receipts, PDFs or screenshots, grab the weights on Hugging Face and try it with vLLM before paying for a separate vision model.