NVIDIA has released Cosmos 3 Edge, a 4-billion-parameter open world model designed to run locally on edge devices. The model enables robots and vision AI agents to understand their surroundings, perform real-time reasoning, and generate robot actions on-device without cloud dependencies (MarkTechPost).
In related infrastructure news, Google is developing “Frozen v2,” a custom server chip that embeds Gemini’s architecture directly into silicon. According to internal sources, the chip could be 6 to 10 times more efficient than current TPUs when deployed, with a target ship date of 2028 (The Decoder). Separately, research from Xiaomi shows that data scale outpaced model size when training robot controllers - performance gains from 100,000 hours of motion capture data substantially exceeded improvements from increasing model parameters, though absolute success rates remain low (The Decoder).