Nemotron 3.5 Bolt E4B 4GB is an open small language model from NVIDIA, released on 23 October 2026 on Hugging Face and build.nvidia.com. It is built on Google's Gemma 4 E4B with NVIDIA's Nemotron post-training, and it is tuned for conversational agents that run on the device: voice assistants, interactive game characters and robotics or IoT interfaces. NVIDIA says it fits that intelligence into as little as about 4GB in its target workloads.
What makes Bolt E4B different
- Built for talking and acting, not coding: NVIDIA argues that a model that excels at agentic coding is often the wrong fit for a fast home assistant.
- Conversational mode: a setting that encourages concise, speech-ready answers.
- Tool use on the device: in NVIDIA's blacksmith demo, a game character checks the player paid, then calls tools to look up a recipe, gather ingredients, craft the item and hand it over.
- Speculative decoding: a DFlash drafter that NVIDIA says gives up to 21% higher throughput than Gemma 4 E4B's original drafter.
Benchmarks and speed
NVIDIA reports roughly 40% higher average scores than its previous Nemotron 3 Nano 4B and about 10% above the other peers it tested, with up to 80% fewer output tokens when reasoning is on. It does not win everywhere: on Tau2-Bench, a multi-turn customer task, Bolt scores 55.1 against 75.1 for Qwen 3.5 4B.
Measured on a Jetson Orin Nano (8GB): about 4.2GB of memory and 15.5 tokens per second, but close to 6 seconds to the first token, so it is not instant on the smallest board. Across NVIDIA's tests, memory use ranged from about 3.1 to 5.0GB.
Where to get it
NVIDIA lists Hugging Face, Ollama, LM Studio, Unsloth, Canonical, Jetson AI Lab and the NVIDIA In-Game Inferencing SDK. Under the hood the model has 8B total parameters (4B effective), quantized to Q4_K_M for the 4GB build.
Our take
If you are building a voice assistant or a game character, pick a model for how it talks and acts, not for its coding score, and test it on your own hardware first. Our course on how LLMs really work is a good primer on why model size, quantization and context change what runs on a device.
