Google debuts Gemini 2.0 Flash to power native multimodal agents

Google just crashed the holiday AI party in a massive way.

While OpenAI was busy dominating headlines with its product countdowns, Google DeepMind unveiled Gemini 2.0 on December 11, setting its sights squarely on the agentic era. The company launched Gemini 2.0 Flash Experimental, a low-latency workhorse that outpaces the previous 1.5 Pro on key benchmarks while running at twice the speed. The entire model was trained and runs exclusively on Google’s sixth-generation Trillium TPUs, which are now generally available to enterprise cloud customers.

Native multimodal output and real-time streaming

Gemini 2.0 Flash shifts the focus from merely understanding multimodal inputs to natively producing them. In a single API call, the model can generate interleaved text, natively generated images, and steerable multilingual audio. All synthetic audio and images come equipped with invisible SynthID watermarks.

Tool use is now native as well. 2.0 Flash can run code, invoke third-party functions, and perform parallel Google Searches to verify information across multiple web sources simultaneously. For real-time applications, Google dropped the Multimodal Live API, which accepts streaming audio and camera video with support for natural voice interruptions. Unlike rival services that lock advanced voice and video behind expensive monthly tiers, developers can test this real-time stream right now for free inside Google AI Studio.

Gemini users on desktop can select the experimental chat-optimized 2.0 Flash right from the model dropdown. Google also launched Deep Research inside Gemini Advanced, which plans multi-step queries across the web and compiles structured research reports with citations.

Agents in the browser and GitHub

Alongside the model, Google previewed three agentic prototypes that take autonomous actions on your behalf.

Project Mariner is an experimental Chrome extension that navigates web pages by reading pixels, text, and form elements. On the WebVoyager benchmark for real-world web tasks, Mariner scored 83.5% as a single agent. To prevent rogue behavior, it is restricted to the active browser tab and requires explicit user confirmation before executing sensitive steps like transactions.

For developers, Jules is a code agent that connects directly into GitHub. It tackles asynchronous Python and JavaScript bug fixes, plans multi-step changes across multiple files, and automatically opens pull requests. Powered by 2.0 Flash and execution tools, the underlying agent system hit 51.8% on SWE-bench Verified.

Finally, Project Astra received an update with near-human latency, a 10-minute conversational memory window, integrations with Google Maps and Lens, and early testing on prototype smart glasses.

Why it matters

Google made a remarkably smart tactical choice by pushing 2.0 Flash out first instead of a slow, monolithic flagship. For real-world AI agents to feel seamless, blazing speed and low latency matter far more than a few decimal points on academic reasoning tests. Slashing latency by half while enabling native audio and visual streaming makes interactive, turn-by-turn copilots genuinely practical.

For developers and power users, the barrier to building dynamic apps just dropped significantly, especially with free access to real-time streaming in Google AI Studio. Rivals who rely heavily on selling premium access to voice and vision models will feel the squeeze.

That said, browser-based automation is still tricky. An 83.5% success rate on WebVoyager is impressive for a benchmark, but in consumer settings, a 16% error rate means transactions fail and forms get mangled. Restricting actions to a single tab with human approval shows Google is cautious, but true autonomy will require far higher reliability before it becomes an everyday habit.