NVIDIA Opens Nemotron 3.5 Super VL, a 120B-Parameter Agent Model Built With Mistral, Perplexity and Thinking Machines Lab

NVIDIA on Thursday released Nemotron 3.5 Super VL, an open-weight multimodal model built for the kind of enterprise agents that never stop running: reading text, images, documents and video, then calling tools, testing code and triaging security alerts inside a single model. The model uses a mixture-of-experts architecture with 120 billion total parameters and only 12 billion active per token, and was distilled from NVIDIA's largest reasoning model, Nemotron 3 Ultra.

The release was built with the Nemotron Coalition — Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs — and ships with open datasets, training recipes and reinforcement learning environments. NVIDIA says pretraining and continued pretraining consumed roughly 26 trillion tokens, with about 3.8 million post-training samples and RL runs across dozens of environments. Third-party listings put the context window at about 1 million tokens, with an introductory price of $0.15 per million input tokens and $0.50 per million output tokens.

The headline claim is throughput. NVIDIA says Super VL delivers up to 5 times the throughput of similarly sized open models on agentic tasks on a single GPU, and matches Nemotron 3 Ultra's accuracy on instruction-following, tool-calling and image-analysis work with about one-fifth the parameters. On the benchmarks the company cites, the model scores 80.8% on CharXiv chart reasoning (84% with Python execution), 63.44% on Hermes Bench versus Ultra's 56.33%, and 87.05% on a LangChain behavioral evaluation versus Ultra's 84.3%. All of these figures are NVIDIA's own.

Inference efficiency gets its own chapter. With NVIDIA's Dynamo software on 16 B200 GPUs, the company says cache-aware routing, session affinity and KV cache offloading let the model serve 8 times as many concurrent users and cut time to first token by 85%. The model runs fully in 4-bit NVFP4 on Blackwell GPUs, and on a DGX Spark desktop NVIDIA recommends its LiLiCorr speculative decoding method, which it says speeds generation by up to 2.3 times. An EXO Labs test on the DGX Spark found no other model that was both more accurate and faster — a result from an independent lab, though one run on NVIDIA's own hardware.

Enterprise pilots are doing the selling. NVIDIA says Scale AI built a supplier-credit processing agent for media company Skydance on Super VL, cutting projected inference cost by 97.5% and latency by 80%; Ramp is using it to convert receipt images into structured JSON with up to 50% higher throughput and 90% lower latency; and AI21 paired it with security firm Alice for threat intelligence and vulnerability analysis, reporting a 40% accuracy gain on those tasks. Each figure is a company-reported projection rather than an audited result.

Two compact companions round out the family. Nemotron 3.5 Bolt 74B-A9B 24GB, a 2-bit-quantized edge model for local agent work on RTX and DGX Spark hardware, is available now alongside Super VL, and a smaller Bolt E4B 4GB model for voice assistants and embedded interfaces is due October 23. Weights are live on Hugging Face, OpenRouter and ModelScope.

The strategic read: NVIDIA is no longer just selling the picks and shovels of the agent economy — it is shipping the models too, open and cheap, tuned to make its own GPUs the path of least resistance. If the 5x throughput claim survives independent benchmarking, the cost of running a 24/7 agent on open weights drops again. If it does not, Super VL still hands enterprises a credible open multimodal option to fine-tune behind their own firewalls.