NVIDIA has introduced Nemotron 3.5 Super VL, a 120B open multimodal agent model. The architecture combines Mamba-2, mixture-of-experts, and attention layers, with each MoE layer containing 512 routed experts. The model supports text, image, and video inputs, with video clips limited to 2 minutes. It was distilled from Nemotron 3 Ultra and trained on 26 trillion tokens.
The model is available via Hugging Face, OpenRouter, ModelScope, and as an NVIDIA NIM. It supports runtimes including vLLM, SGLang, and TensorRT-LLM on Blackwell and Hopper hardware. For production, Dynamo recipes enable KV-aware routing, which NVIDIA claims can support 8x more users on 16 B200 GPUs while reducing first-token latency by 85%.
NVIDIA also released the Nemotron 3.5 Bolt 74B-A9B, a distilled version of Super VL designed for local agents. It features native 2-bit quantization, allowing it to fit into 24GB of memory on RTX and DGX Spark systems. Additionally, the company announced the Bolt E4B 4GB model, scheduled for release on October 23, aimed at voice assistants and embedded agents.
