AI news · October 17, 2026
NVIDIA opens Nemotron 3.5 Super VL for multimodal agents
Nemotron 3.5 Super VL is open now with 120B total and 12B active parameters; NVIDIA claims up to 5× throughput, so a local or hosted trial can test cost against quality.
- total parameters
- 120B
- active parameters per request
- 12B
- reported throughput gain
- 5×
NVIDIA announced Nemotron 3.5 Super VL on October 15 as an open model for agents that work across text, images, documents, and video. The model has 120 billion total parameters and activates 12 billion for each request. NVIDIA says the release includes open weights, training data, and code, with distribution through Hugging Face, OpenRouter, ModelScope, and NIM.
The company reports up to five times more throughput than comparable open models on one GPU, but the number depends on hardware, batch size, and task. The smaller Nemotron 3.5 Bolt models are aimed at local use in 24GB or 4GB memory footprints; NVIDIA lists llama.cpp, Ollama, LM Studio, Unsloth, and vLLM support. These are strong access points, not an independent performance audit.
What you can do with it
Start with one multimodal task that matters: extract fields from documents, inspect screenshots, or find events in short clips. Run the same set on your current model and Nemotron, then record memory, time, tool-call errors, review time, and cost before moving to a larger batch.
Our take
Open weights plus a sparse active count make this more interesting than a bigger closed model announcement. The practical gain depends on serving quality and memory, not the 120B headline. A short local comparison can show if the model saves money without lowering accepted results.
Source: NVIDIA Blog — Made With Models writes the brief; the reporting is theirs.