NVIDIA has launched Nemotron 3.5 Super VL, an open 120-billion-parameter vision-language model tailored for always-on enterprise agents. The model activates 12 billion parameters per token using an architecture that interleaves Mamba-2, mixture-of-experts, and attention layers to handle text, image, document, and video inputs in a single checkpoint.

According to NVIDIA, Super VL outperforms its larger predecessor, Nemotron 3 Ultra, across key agent benchmarks, scoring 63.44% on Hermes Bench and 87.05% on LangChain evaluations. The company also released open datasets, training recipes, and reinforcement learning environments in NeMo Gym alongside two smaller edge models under the OpenMDW-1.1 license.

The model is available on Hugging Face, OpenRouter, ModelScope, and as an NVIDIA NIM, supported by vLLM, SGLang, and TensorRT-LLM runtimes on Hopper and Blackwell hardware. NVIDIA reports that production deployments using Dynamo recipes can deliver up to 8x higher user capacity on B200 GPU setups.

Why it matters

  • Consolidates text and vision capabilities into a single open model, simplifying agent deployment architectures.

  • Reduces infrastructure overhead with high throughput and low first-token latency on modern GPU hardware.

  • Provides open datasets and RL environments to help teams fine-tune custom enterprise agents.

Source: marktechpost.com