According to CoinMeta, NVIDIA has open-sourced Bolt E4B, which is a device with 8GB of memory that is capable of running both visual and dialogue models simultaneously. In addition, NVIDIA has also introduced a smaller AI model, Nemotron, with 3.5GB of memory, specifically designed for local voice assistants, game characters, and robots. This model is further trained based on Google's Gemma and E4B, enhancing its capabilities in continuous dialogue, command understanding, and tool invocation. The model has a total of about 8 billion parameters, but a large portion of these parameters are used simply by looking up tables, with the core computational scale being around 4 billion parameters. NVIDIA also showcased desktop robots Reachy Mini and Cosmos; the latter two models have 3GB of memory each. Edge is responsible for recognizing the surrounding environment, while Bolt handles chatting and invokes preset programs to control the robot's actions. Both models run simultaneously on a single Jetson Orin Nano with only 8GB of memory.