Meituan’s Tabbit AI browser brings multi‑model agents to the desktop

Tabbit is an important proof point that the browser is becoming a primary interface for agentic AI. Instead of locking users into a single model, Meituan’s GN06 team is wiring multiple Chinese frontier models behind an agent layer that can orchestrate complex cross‑site workflows. That’s strategically significant in a market where DeepSeek, Zhipu GLM, Kimi and others are all vying to be the default system prompt for Chinese netizens: Tabbit shifts power toward whoever controls the orchestration layer rather than just the underlying model APIs.

From an AGI‑race perspective, Tabbit shows how quickly agent‑centric design is becoming mainstream. An AI‑native browser that can call dozens of models to complete tasks on other websites starts to look more like a lightweight operating system than a traditional browser. That architecture will generate rich telemetry about what agents can and can’t do reliably in the wild—data that can feed back into both model training and safety work. It also normalises the idea that day‑to‑day computing is mediated by model‑driven agents rather than direct user clicks.

For the Chinese ecosystem, Meituan’s move both hedges and amplifies. It hedges by staying model‑agnostic, but amplifies competition by making it trivial for users to route workloads to whichever model performs best. That, in turn, keeps pressure on local labs to iterate quickly on coding, planning and tool‑use capabilities.