NVIDIA brings Bolt to local agent workflows

NVIDIA says it designed Bolt for everyday agentic tasks on local systems, including RTX and DGX Spark machines.[2] It says users can delegate web research, writing and code-debugging projects through local agent harnesses without sending the work to the cloud.[2] The stated use case is therefore not limited to standalone prompts: NVIDIA describes Bolt as a model that can be connected to agent software to handle projects.[2] Teams evaluating it should distinguish that intended workflow from evidence that a particular agent setup will perform well on their own tasks.

NVIDIA says native 2-bit quantization packs Bolt into 24GB of memory, and that the model can run alongside other applications while using that packed model.[2] The company says Bolt is available on Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.[2] NVIDIA also lists GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations as systems on which Bolt runs.[2] RTX Spark laptops built for Bolt are due to arrive on October 16, according to NVIDIA.[2]

NVIDIA’s model card lists text, images and video as Bolt’s inputs, with text as its output.[3] It lists a context length of up to 256,000 tokens.[3] NVIDIA describes Bolt as open source, including its training data, weights and recipes, and the model card says it is ready for commercial or non-commercial use.[4][3] Those details give teams several concrete points to check when assessing fit: whether their harness supports the available tooling, whether their work fits the listed modalities and context limit, and whether the published terms meet their needs.

NVIDIA says Bolt aims to balance speed and intelligence per gigabyte, and claims it is faster than popular dense models while being smarter than popular MoE models with similar memory footprints.[3] NVIDIA ranks Bolt as the top open model in its class on PinchBench and second among open models on BrowseComp.[2] Separately, Artificial Analysis measured Bolt at 21.4 on its Intelligence Index, placing it fourth among models with fewer than 100 billion parameters.[5] Steel said it partnered with NVIDIA to put Bolt on its leaderboard from launch day, and described it as a best-in-class open model on BrowseComp and WebArena.[6] These claims and measurements use different evaluations; they do not establish how Bolt will perform on a team’s specific workload.