NVIDIA released Nemotron 3.5 Bolt E4B 4GB on October 23, giving developers an open-weight model aimed at conversational agents running locally. Its focus is practical: assistants that can follow a conversation, call application tools and respond without relying on a cloud model for every turn.
A small model for dialogue and actions
Built on Gemma 4 E4B, the model combines:
- Multi-turn tool calling for actions inside applications.
- Configurable reasoning and a conversational mode for shorter, speech-friendly replies.
- A 64K context window in NVIDIA’s published Q4_K_M model.
The checkpoint is text-in/text-out: a voice assistant still needs separate speech components. Weights and deployment examples are available on Hugging Face. NVIDIA’s October 15 family announcement previewed Bolt E4B; the October 23 release is the availability milestone.
The memory budget still depends on the workload
The “4GB” name is a target, not a universal memory guarantee. NVIDIA’s tests range from about 3.1GB to 5GB depending on hardware and workload. Context, runtime configuration and the rest of an application still matter when evaluating a device’s memory budget.
NVIDIA also reports roughly 40% average benchmark improvement over Nemotron 3 Nano 4B. That is a vendor-reported aggregate, not an independent evaluation or a promise that every application will improve by the same amount.
Why it matters for builders
Useful agents do not always need the biggest model. For builders such as SunMarc App Labs, this is worth evaluating for bounded desktop companion workflows—finding a setting, explaining a feature or invoking a small set of app actions.
A practical trial would check correct task completion, response time and peak memory on the intended device. This is an evaluation opportunity, not an announced SunMarc assistant feature or a claim that we have tested the model.
Our coverage of Recursive Flint for local coding agents explores a different local-model workload. Here, the useful test is conversational task accuracy within the device’s memory and response-time budget—not a general chatbot leaderboard.