A new arXiv paper introduces Nemotron 3.5 Bolt 74B-A9B 24GB, a vision-language model designed for local AI tasks such as using tools, browsing the web and working in a terminal. Its authors say it is built for memory-constrained computers and reports 84 TPS on RTX Spark, a measure of inference speed. The abstract does not give the test settings behind that speed figure.
The model combines Mamba and Transformer designs, two approaches to building AI models. It is derived from the Nemotron 3.5 Super VL 120B-A12B family and uses Star Elastic compression and 2-bit Quantization-Aware Distillation. The authors present those methods as a way to balance performance with memory use. The model’s name includes “24GB,” but the abstract does not spell out hardware requirements or say whether that is the minimum memory needed.
What tasks is it designed for?
The paper focuses on agentic tasks: work where a model uses tools or interacts with software to carry out a goal. The authors say they optimized Bolt for reasoning, tool use, web browsing and terminal interactions. They also trained a companion system called LiLiCorr, described as a speculative drafter, and say it is released alongside the model to help reach the target inference speed. The abstract does not explain how to install either system or what license applies.
The authors report higher accuracy than leading models with similar memory footprints and speed across benchmarks including GDPVal v2, BrowseComp and Terminal-Bench 2.1. They also say Bolt’s speed is comparable to models with much lower active parameter counts and is sufficient for long-running agent tasks. The abstract provides no benchmark scores or model-by-model comparisons, so readers cannot assess the size of those gains from the summary alone.
What can users conclude?
The paper positions Bolt as an option for people who want a vision-and-language model to handle software tasks on local hardware, rather than relying on a remote service. Its reported 24GB target and RTX Spark speed figure offer a sense of the intended deployment, but they do not establish performance on other computers. The paper’s claims are from the model’s authors; the abstract does not provide independent test results or access details.
Why it matters
For people hoping to run an AI agent on their own computer, Bolt is presented as a model aimed at fitting local tasks into a 24GB memory footprint. The authors report promising speed and benchmark results, but the abstract does not say which hardware beyond RTX Spark can run it or provide detailed scores, so practical performance remains uncertain.
