NVIDIA released Nemotron 3.5 Bolt E4B on October 23, adding a compact conversational model to its push for AI agents that run on local hardware. The release follows the larger Super VL and Bolt 74B models announced earlier this month. NVIDIA’s release record confirms availability on its developer platform and Hugging Face.
The practical pitch is a smaller model that can hold a conversation and invoke application tools without depending entirely on a remote, large-model service. Whether it delivers that experience depends on the application and hardware, not just the model’s name.
What the new release contains
The downloadable 4GB model card describes a quantized, text-input and text-output model built from Google’s Gemma 4 E4B and further trained by NVIDIA. It lists eight billion total parameters, four billion effective parameters, and a context window of up to 64,000 tokens.
Developers can switch reasoning on or off and use a conversational mode aimed at shorter responses. A separate DFlash companion model supports speculative decoding. NVIDIA also offers a full-precision BF16 version, which its developer model card positions primarily for customization and research rather than direct production inference.
The repository identifies the release’s governing license as OpenMDW 1.1. Teams should check those terms for their intended use rather than assume every model in the wider Nemotron family has the same license.
Why a smaller agent model matters
In its release explanation, NVIDIA targets device assistants, interactive characters and embedded interfaces. It describes a footprint as low as roughly 4GB in target workloads.
That figure should be treated as a deployment target, not a promise that a complete assistant will fit into any 4GB device. Applications also need room for conversation state, their own software and, where relevant, speech or vision components. Bolt E4B’s text interface does not by itself supply an end-to-end voice assistant.
For developers, the appeal is more control over where inference happens and how a model connects to local applications. The trade-off is taking responsibility for integration, resource use and safeguards around actions. Local execution alone does not establish that an agent is accurate, private or safe.
Claims that still need application testing
NVIDIA reports improved average conversational benchmark results against its earlier Nano 4B model and shorter outputs in selected comparisons. Those are vendor measurements. Its published tables also show that Bolt E4B does not lead every comparison, and the research post says two new evaluation suites will be released later.
PromptsZone has not benchmarked the model. A useful next test is whether it follows instructions, chooses the right tools and responds quickly enough in a real application, including when requests are ambiguous or a tool fails. Aggregate scores cannot answer that for an individual deployment.
Photo: Coolcaesar, NVIDIA headquarters, August 2018, via Wikimedia Commons, CC BY-SA 4.0. Commons version cropped and perspective-corrected by Jacek Halicki; displayed at a reduced size. Archival image, not the launch event.
