NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model

MarkTechPost reports NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal mixture-of-experts model with 12 billion active parameters, a 1-million-token context window, and NVFP4 weights. The piece frames it as an open multimodal agent model in NVIDIA’s Nemotron line.

Sparse multimodal agents with million-token context are becoming the default open release shape — Meta’s Muse Spark Open in the same week, NVIDIA’s Super VL here. The useful distinction is deployment: NVIDIA is explicitly wiring Super VL into NIM, NeMo, and its desktop-to-data-center hardware story.

We’ll keep model-card numbers tied to NVIDIA and MarkTechPost’s reporting until outside benches land. For teams already on CUDA, an open VL agent model you can post-train is practical progress. I’m watching whether 12B-active MoE delivers the latency NVIDIA advertises under real tool-use loads.

Companion NVIDIA blog coverage adds Bolt edge models, training token counts, and named enterprise customizers.