LONDON — Google DeepMind today unveiled Gemini 3.0, introducing a unified multimodal neural network trained natively from the ground up on interleaved audio, high-resolution video streams, tactile sensor inputs, and code. The flagship architecture achieves human-parity turn-taking latency during live conversational exchanges while expanding the continuous active context window to 10 million tokens.
The release marks a significant milestone in DeepMind’s quest to build proactive, ambient cognitive companions capable of seeing, hearing, and navigating physical environments alongside human users.
Native Sensory Interleaving
Unlike previous multimodal approaches that cascaded separate transcription, text processing, and speech synthesis layers, Gemini 3.0 natively ingests raw audio frequencies and high-frame-rate video signals within a single transformer backbone. This allows the model to detect subtle vocal inflections, emotional hesitation, background acoustic nuances, and rapid physical gestures with unprecedented fidelity.
In live demonstrations presented at Google’s Kings Cross headquarters, Gemini 3.0 observed a complex mechanical watch repair through a pair of lightweight camera glasses, offering real-time verbal guidance and correcting tiny gear misalignments before the technician completed the maneuver.
Long-Horizon Context Memory
The 10-million token context window allows software developers to pass entire multi-gigabyte operating system codebases, architectural documentation, and decades of git history into a single active session. DeepMind executives confirmed that developer API access is opening today via Google Cloud Vertex AI.