NVIDIA releases Nemotron 3.5 Bolt E4B

NVIDIA describes Bolt E4B as a small language model optimized for speed and intelligence per gigabyte in conversational agents.[1] The company says it builds on Gemma 4 E4B with Nemotron post-training, combining natural conversation with multi-turn agentic capabilities.[1] NVIDIA says the model is intended for voice and device assistants, interactive characters, and robotic or IoT interfaces.[1] Those use cases place conversation and task completion in the same model: NVIDIA says Bolt E4B can call application APIs to carry out tasks across a conversation.[1]

NVIDIA says a conversational mode can encourage concise, speech-ready responses for back-and-forth interactions.[1] Users can enable reasoning for complex requests or disable it for faster responses, the company says.[1] The model has a 64,000-token context window, which NVIDIA says can hold conversation history, instructions, tool definitions, retrieved information and application state.[1] For teams evaluating it, these are configurable behaviors and stated capacity; the evidence does not establish how the model performs on a particular organization’s tasks or systems.[1]

NVIDIA reports that Bolt E4B has roughly 40% higher average scores than its prior model, Nano 4B, and scores roughly 10% higher than other evaluated peers.[1] The company also says the included DFlash parallel drafter delivers up to 21% higher throughput than Gemma 4 E4B’s original MTP drafter.[1] These figures offer comparison points for evaluation, but the evidence does not provide benchmark names, test conditions or results for a specific deployment. Teams should treat them as NVIDIA’s claims and test their own workloads before relying on the comparisons.[1]

NVIDIA announced Bolt E4B’s availability on Hugging Face, and its research page lists Hugging Face, Ollama, LM Studio, Unsloth, Canonical, Jetson AI Lab and the NVIDIA In-Game Inferencing SDK as ways to get started.[2][1] NVIDIA’s post describes the release as open and customizable, including weights, data and recipes.[2] The model card says it is ready for commercial or non-commercial use.[4] NVIDIA’s build page also offers a free API endpoint for building.[3] The available evidence does not specify other pricing or deployment terms.[3][4]