NVIDIA Reveals Nemotron 3.5 Super VL and Bolt: What to Check First for Local Agents

NVIDIA unveiled two open models, Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, on Oct 15th 2026 (U.S. time). The key point is that they form one family spanning large agent models running in data centers to small models running on PCs and edge devices. This article focuses on what to actually check when using them locally.

1. Dividing the roles of the three models

Super VL is a 120B (120 billion parameters) open model for agents that understand images, web pages, and PDFs together. It is designed to read tables, charts, and text in documents and handle multi-step workflows. NVIDIA explains that it delivers accuracy similar to the largest reasoning models with about one-fifth the parameter size.

Bolt 74B-A9B 24GB is for agents running locally on PCs. It is fitted into 24GB of memory through native 2-bit quantization and is presented as capable of handling tasks such as web research, writing, and code debugging without sending them to the cloud. Bolt E4B 4GB is intended for uses where fast responses are important, such as voice assistants and conversational characters, and is scheduled for release on October 23.

2. Three things to check first when using them locally

First, available memory. 24GB is the figure for the model weights. Actual usage may be higher when handling long contexts or running a browser and development tools together, so you should check it directly using the length of your own tasks.

Second, the quality of 2-bit quantization. The presentation materials alone cannot show how much accuracy declines in exchange for reduced memory use. It is safer to compare it with existing models under the same conditions using several document-summarization or code-editing tasks you frequently perform.

Third, execution tools. Bolt 24GB is stated to be usable with llama.cpp, Ollama, LM Studio, and vLLM. Quantization files and default settings may differ by tool, so check the model card’s recommended settings and license terms before use.

3. Read benchmark figures as company claims

NVIDIA stated that Bolt 74B-A9B 24GB ranked first among open models in the same tier on PinchBench and second on BrowseComp. It explains that Super VL delivers up to five times the throughput of models in the same tier on the Artificial Analysis Intelligence Index. Since both figures were presented by NVIDIA, they should be considered alongside independent benchmark results and your own task results.

4. What changes for enterprise adoption

The open weights, training data, and training recipe are being released together, allowing companies to perform post-training with their own data and operate the models on their own infrastructure. NVIDIA also stated that serving Super VL with NVIDIA Dynamo can handle up to eight times more users and reduce latency by 85%. This is also a company-provided figure.

NVIDIA introduced Atlassian, CrowdStrike, Perplexity, and Salesforce as integration partners. However, the actual adoption results of these companies cannot be confirmed from the announcement alone. If you are considering adoption, evaluation-set results using your own data matter more than the partner list.

Conclusion

Super VL is an option for server-side agents, while Bolt is an option for PC-side agents. This is not a reason to change your existing work environment immediately, but for people who want to run agents locally, it establishes a new comparison point: an open model in the 24GB class. For the RTX Spark laptop story targeting Bolt execution, the previously published RTX Spark article ↗ may also be helpful.

Source: NVIDIA blog, 'Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models' (announced Oct 15th 2026) Read the original ↗. The performance figures are based on NVIDIA’s announcement and are not independently verified.

This English edition was written and translated with AI assistance, with sources provided. For errors or additional information, please request a correction. Editorial policy