NVIDIA releases Bolt for local conversational agents

NVIDIA’s October 23 announcement introduces Nemotron 3.5 Bolt E4B, built on Gemma 4 E4B with additional NVIDIA training. Its intended applications need conversation interleaved with actions: an assistant must decide when to answer, ask for clarification or call an application function. NVIDIA cautions that benchmarks cover only part of these requirements; improvements must translate into better behavior in the actual application.

The model card specifies text input and output. A conversational setting encourages shorter, less formatted replies and can be combined with reasoning and tool use. Delayed tool results require a particular message sequence: immediately acknowledge the call with a tool message marking its result pending. Conversation can then continue. Once the result arrives, provide it as structured information in a user message. The template does not accept an unrelated tool response inserted later; the application must preserve this distinction.

For developers trying local text chat, the Jetson guide describes a Docker and llama.cpp setup. Selecting the hardware module determines the container; the model download is cached for later runs. The guide places opening the chat interface or using its API after the server becomes ready. Its approximately four-gigabyte figure concerns the model’s target memory footprint: total memory consumption also depends on conversation length and runtime overhead.

MarkTechPost also covered the release, linking readers to NVIDIA’s underlying materials.