Google DeepMind debuts Gemini 2.0 Flash and advances Project Astra with real-time multimodal perception
Google DeepMind has unveiled Gemini 2.0 Flash alongside major capability upgrades for Project Astra, delivering sub-second multimodal audio/video streaming, 10-minute visual memory, and real-time agentic tool use.
Google DeepMind has taken a decisive step into the 'Agentic Era' with the simultaneous release of Gemini 2.0 Flash and an advanced iteration of its Project Astra universal assistant prototype. Engineered from the ground up for ultra-low latency, Gemini 2.0 Flash runs at more than twice the inference speed of Gemini 1.5 Flash while integrating native multimodal output for both speech synthesis and image generation directly inside the model architecture.
The architectural improvements power Project Astra's conversational leap. In real-world testing, Astra demonstrated human-speed conversational cadence with seamless multi-language code-switching and up to ten minutes of persistent in-session visual memory. The assistant can continuously monitor video camera streams, track physical objects in dynamic environments, and recall environmental changes that occurred minutes earlier.
Alongside Project Astra, Google showcased Project Mariner—an experimental browser-agent operating inside Chrome that utilizes pixel-level visual understanding to navigate web forms, compare product listings, and automate research tasks. Gemini 2.0 Flash is available immediately to developers via Google AI Studio and Vertex AI, with rollout expanding into Gemini Live on mobile and Android XR ecosystems.
Why it matters
With native real-time audio and visual generation running at double the speed of Gemini 1.5, Gemini 2.0 Flash establishes the foundation for always-on embodied assistants, smart eyewear, and browser automation agents.