Physix FrontierPhysix Frontier · News Briefing Card (IT Home · Oct 15, 2026)
NVIDIA Unveils 120B-Parameter Multimodal Model Nemotron 3.5 Super VL
KEY FACTS
- NVIDIA launched the Nemotron 3.5 Super VL multimodal model, with 120 billion total parameters and 12 billion active parameters.
- The model delivers up to 5x the single-GPU agent task throughput of open models at the same scale.
- The model integrates reasoning, tool calling, and text-image-video understanding, targeting multi-step agent tasks.
- NVIDIA also released the Bolt 74B-A9B 24GB edge model, compressed to 24GB via 2-bit quantization.
- Bolt E4B 4GB is planned for launch on October 23, while Super VL and the 74B version are already available.
KEY DATA
120 billionTotal Parameters
12 billionActive Parameters
up to 5xThroughput Gain
80.8%CharXiv Score
PHYSIX OBSERVATION
NVIDIA is pushing the large-model race toward agent workloads: MoE plus quantization multiplies single-GPU throughput several times over, which directly lowers the cost for enterprises running multi-step tasks. Opening the weights and training recipe is meant to lock users into the NeMo ecosystem against other open-source camps. The near-term benefit for users is a lower barrier to local deployment, but real gains still depend on post-training adaptation.
Source: IT Home report