Nvidia open-sources Bolt E4B: 8GB memory devices can simultaneously run vision and dialogue models.

Beating AI News Flash: NVIDIA has open-sourced a small AI model, Nemotron 3.5 Bolt E4B 4GB, specifically designed for local voice assistants, game characters, and robots. It is continuously trained based on Google's Gemma 4 E4B, enhancing capabilities in continuous dialogue, instruction understanding, and tool calling. The model has approximately 8 billion total parameters, but many parameters only require table lookups and do not need complex computation, with the core computational scale at about 4 billion parameters.

NVIDIA also showcased the desktop robot Reachy Mini. Cosmos 3 Edge is responsible for recognizing the surrounding environment, while Bolt handles chatting and calls preset programs to control the robot's movements. Both models run simultaneously on a Jetson Orin Nano with only 8GB of memory.