Gemma 4 Open-Weight Models Run Frontier AI on Single GPU; Gemini 3.1 Flash Live Debuts with Voice and Cross-Platform Import

Google DeepMind released Gemma 4, a family of four open-weight multimodal models that fit on a single 80GB Nvidia H100 GPU while delivering frontier-competitive performance; the family includes Gemma's first Mixture-of-Experts model, supports over 140 languages, and covers image, text, and audio inputs under Apache 2.0 license with vLLM and llama.cpp optimization. Separately, Google released Gemini 3.1 Flash Live, a voice-first model optimized for real-time natural-rhythm conversations with background noise filtering, cross-platform chat import from ChatGPT and Claude to reduce switching friction, free Personal Intelligence integrating Gmail, Calendar, Drive, and YouTube, and persistent memory for conversational continuity. Google DeepMind also formed a research partnership with Agile Robotics to combine Gemini Robotics foundation models with real-world robotic operation data for industrial AI robotics.