NVIDIA launches Nemotron 3.5 model family

Nvidia has expanded its Nemotron 3.5 family with Nemotron 3.5 Super VL, a 120B-parameter vision-language model aimed at AI agent workloads, alongside two lightweight models built for edge devices. The company frames the release as rounding out its deployment stack from data centers down to endpoints.

NVIDIA has introduced Nemotron 3.5 Super VL, an open multimodal model built for AI agents that run continuously and handle multi-step tasks. It uses a MoE architecture with 120B total and 12B active parameters; NVIDIA claims up to 5x the throughput of comparable open models on a single GPU for agent workloads, and combines text, image, document and video understanding with reasoning, instruction following and tool calling. Developers can post-train it with their own data via the NeMo toolchain.

According to a Hugging Face page shared on Reddit's LocalLLaMA, NVIDIA has published Nemotron 3.5 Bolt 74B-A9B 24GB, a vision-language model aimed at everyday agentic tasks running locally. It uses native 2-bit quantization to fit in 24GB of memory and run alongside other applications, targeting the best intelligence and speed per GB, with GGUF builds already posted.

NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter model it says delivers the fastest task completion and highest accuracy for always-on agents. Two edge models join it: Nemotron 3.5 Bolt 74B-A9B 24GB and Bolt E4B 4GB, tuned to balance intelligence and speed per GB for local agentic workloads.

In a developer blog post, NVIDIA explains how Nemotron 3.5 Super VL fits into always-on agents that take on longer enterprise workflows, where a single request can mean reading a document and checking information across sources.

A post on Reddit's LocalLLaMA community says NVIDIA's Nemotron-3.5-Super-VL-120B-A12B model family has been released, covering base, NVFP4, and BF16 builds plus LiLiCorr speculative decoding. The thread offers no benchmark numbers or license details beyond the release notice.