Google Launches Gemini 3.8 and Sends TPUs Into Orbit

Google's latest release introduces Gemini 3.8 Flash TTS and Flash-Lite TTS, which generate custom voices from natural-language prompts across more than 100 languages. Instead of relying on presets, developers can describe a character or replicate a voice using 30 seconds of verified audio. The models support inline directions like laughs or sighs, and stage two-speaker conversations. Flash-Lite targets high-volume tasks, while Flash is built for detailed creative direction. On Hume AI's Voice Design Benchmark, Google reported a top score of 71.4. Output is secured with SynthID watermarks and C2PA credentials, though voice replication is restricted in regions like the UK, EEA, India, Texas, and Illinois.

For enterprise applications, Gemini 3.8 Live with Live Avatar enables real-time video agents with synchronized lip-syncing in 97 languages. Operating with a 128,000-token input context window and a 24,000-token output limit, the system supports asynchronous tool calling so agents can fetch data mid-conversation. Meanwhile, Gemini Notebook adds Interactive Learning Overviews with quizzes and 60-second Video Overviews in over 80 languages, alongside a hands-free Live Chat feature on mobile supporting nearly 100 languages.

Looking to the future of infrastructure, Google launched Project Suncatcher. Partnering with Planet, the company sent four TPUs into orbit aboard a refrigerator-sized satellite named MVP, launched via a SpaceX Falcon 9 rocket from Vandenberg Space Force Base on the Transporter-18 mission. The satellite draws 1 kilowatt of solar power to run Gemini queries in 15-minute bursts. The one-year mission tests whether the hardware can withstand 10g of launch acceleration, transient forces between 50g and 100g, passive vacuum cooling, and radiation. Google envisions a future 81-satellite cluster operating within a 1-kilometer radius, with a two-satellite laser-link test planned for 2027.

For developers, these updates shift AI interaction from static text to highly directable, real-time multimodal experiences. The ability to program vocal nuances and run background tools during live avatar chats allows for highly sophisticated customer service agents. Simultaneously, the orbital tests signal a long-term shift toward space-based computing that could bypass terrestrial energy constraints by capturing up to eight times more solar power.