Nvidia’s 120B Super VL agent model has open weights and costs $0.10 per million tokens

Nvidia has released three open models for AI agents: Nemotron 3.5 Super VL, a 120-billion-parameter vision-language model for data centers, and two Bolt models that run locally on PCs. Nvidia says Super VL is five times faster than the next best model in its class, as TechCrunch reported.

Three models for three kinds of hardware

In its announcement blog post, dated October 15, Nvidia describes Super VL as a model for always-on agents that combines reasoning, tool use and image understanding. It says the model approaches the accuracy of its largest reasoning model at roughly one-fifth of the parameter count. Super VL was pretrained on about 26 trillion tokens, then post-trained on roughly 3.8 million samples with reinforcement learning across dozens of environments, according to Nvidia.

The Bolt family targets local use. Nemotron 3.5 Bolt 74B-A9B 24GB uses native 2-bit quantization so that it fits in 24GB of memory, and is meant for multistep reasoning, tool calling and web browsing on a PC. Nvidia says it ranks first among open models in its class on PinchBench and second on BrowseComp. A smaller Bolt E4B 4GB is aimed at conversational and physical agents, including on Jetson devices.

Where to download them and what runs them

Super VL is available now on Hugging Face, OpenRouter and ModelScope, and as an Nvidia NIM microservice. Bolt 74B-A9B is on Hugging Face and works with llama.cpp, Ollama, LM Studio, Unsloth and vLLM. Bolt E4B follows on October 23.

Nvidia says Super VL runs on DGX Spark and DGX Station, RTX PRO and RTX systems, Jetson, and in cloud and data center environments. Bolt 74B-A9B is built for the RTX Spark laptops that Nvidia says arrive October 16, and also runs on GeForce RTX PCs and DGX systems. Nvidia lists Atlassian, CrowdStrike, Dassault Systèmes, Perplexity, Salesforce and Synopsys among software partners, and Scale AI, AI21, Mercor and Ramp among those customizing the model. TechCrunch adds that Skydance and EY are also using it.

The blog describes open weights, open datasets and training recipes, with reinforcement learning environments in NeMo Gym. The post, as retrieved, does not name a license, so commercial terms should be checked on the model cards.

A cheap model with few active parameters

The most useful context comes from market data rather than Nvidia’s own claims. According to our roundup of this week’s models, a pricing tracker lists Super VL at about 12 billion active parameters per token, with a context window of 1,048,576 tokens. It is priced at $0.10 per million input tokens and $0.40 per million output tokens, the cheapest model on that tracker.

For comparison, the same roundup lists DeepSeek V4.1-Flash at $0.15 and $0.60 per million tokens under an MIT license. It lists Meta’s open-weight Muse Spark Open at about 816 billion total parameters with 42 billion active. Super VL is therefore far smaller than Meta’s open model while using about a quarter of its active parameters. Meta’s more capable Muse Wonder models are built for agents too, but they are served through Meta’s own API and not as open weights.

That points to the real trade-off. A small active-parameter count makes a model cheap to serve, and Nvidia claims 85% lower latency and up to eight times more users served with its Dynamo software. The company does not say what hardware or workload those figures come from.

What Nvidia has not shown

Nearly every performance number here is Nvidia’s own. TechCrunch notes that Nvidia says Super VL finishes post-trained agentic tasks faster than models from Anthropic and OpenAI, on work such as document intelligence, financial accounting and medical insurance claims. Those comparisons depend on the model being post-trained for each task, and the blog gives no full benchmark table in the text we could read. The Bolt rankings cover only a class that Nvidia defines.

The strategy is clearer than the benchmarks. The company’s biggest customers, the proprietary labs, are also turning to other chips such as AWS’s Trainium and Cerebras, as TechCrunch notes. Open models that run best on Nvidia hardware give enterprises a reason to stay on its stack. Briski put the pitch this way:

Unlike closed frontier models, organizations can post-train Super VL on their own data

For consumers, the pitch is privacy: according to the TechCrunch report, Nvidia says PC users can hand off everyday projects “without sending those workloads to the cloud.” Whether a 74-billion-parameter model squeezed into 24GB handles real browsing and tool use reliably is untested here, and independent evaluations will be the first check on that claim.