Nvidia has released the Nemotron 3.5 Bolt E4B, a 4GB model designed for local voice assistants and AI agents. The model launched on October 23, 2026, as part of the Nemotron 3.5 family, which focuses on on-device processing rather than cloud-based solutions.
The Bolt E4B is a post-trained version of Google's Gemma 4 E4B model. It features 8 billion total parameters, 4 billion effective parameters, and supports a 64K context window. The model is optimized for hardware with limited memory, specifically GeForce RTX GPUs and Jetson edge devices.
Nvidia reports that the Bolt E4B achieves approximately 40% higher benchmark scores compared to the previous Nemotron 3 Nano 4B. Additionally, the model is capable of reducing output tokens by up to 80% in certain configurations.
The model is fine-tuned for multi-turn tool use and supports asynchronous tool calling, allowing agents to perform tasks without interrupting conversations. It supports 35 languages and is intended for use in robotics, interactive characters, and voice assistants.
Nvidia also released two larger models in the Nemotron 3.5 series on October 15: the 120 billion parameter Nemotron 3.5 Super VL and the 24GB Nemotron 3.5 Bolt 74B-A9B. The weights and training data for these models are available under the OpenMDW-1.1 license.
