Nvidia releases 4GB Nemotron 3.5 Bolt for local AI agents

Nvidia released the open-source lightweight language model Nemotron 3.5 Bolt E4B 4GB on October 23, 2026, according to aitimes.com. The model is designed to run in environments with roughly 4GB of memory, targeting local agents, games and robots.

Nemotron 3.5 Bolt is built on Google's Gemma 4 E4B, with Nvidia applying its own post-training techniques. It runs on Nvidia hardware including GeForce RTX, Jetson, RTX Spark and DGX Spark, and supports a context length of up to 64,000 tokens.

On main conversational agent benchmarks, Nvidia says the model scored about 40% higher on average than the earlier Nemotron 3 Nano 4B and about 10% higher on average than comparison models. With reasoning enabled, it cut output token counts by up to 80%. It beat competing models on 15 of 21 benchmarks, though Qwen3.5-4B led on some long-context and tool-calling evaluations, scoring 75.1 on Tau2-Bench against 55.1 for Bolt.

The model can call APIs during a conversation to carry out real tasks and supports multi-turn task continuation. It offers a conversational mode for short answers suited to voice interfaces, and reasoning mode can be switched on or off. It supports 35 languages, while the base model was pretrained on more than 140.

Training used RLVR and RLHF with the Nemo-RL and Nemo-Gym frameworks. Post-training covered tool calling, instruction following, structured output and multi-turn dialogue, with non-reasoning and reasoning data kept at a 2:1 ratio. The released model was quantized to Q4_K_M format from a selected BF16 checkpoint, with imatrix calibration using about 1 million tokens of reinforcement learning data.

Availability and extra tooling

Nvidia also released a DFlash drafter for parallel inference that raises throughput by up to 21% compared with Gemma 4 E4B's existing drafter. The model is available through Hugging Face, Ollama, LM Studio, Unsloth and Jetson AI Lab, and can be applied to game characters using Nvidia's in-game inference SDK.