Artificial Intelligence

Meituan Launches CatPaw, a Full-Scenario AI Agent Platform Built on a 1.6T-Parameter Model

By 美团技术团队 · Jul 28, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

CatPaw is one of the first large-scale agent platforms backed by a trillion-parameter model trained on non-Nvidia silicon, and it arrives with internal proof of 90,000 users. For teams evaluating agent infrastructure, it offers a reference architecture that spans personal productivity and enterprise deployment with baked-in domain expertise rather than a blank-slate LLM.

Summary

Meituan has released CatPaw, an AI agent platform powered by its LongCat 2.0 model—a 1.6-trillion-parameter system trained entirely on domestic hardware. The platform splits into two parts: a personal AI workbench that runs across mobile, desktop, and cloud, and a managed agent service for building digital employees inside tools like Feishu and WeCom. Agents can autonomously execute tasks like file manipulation, browser control, and terminal commands, delivering finished reports or code.

CatPaw layers Meituan's decade-plus of local commerce data on top of general agent capabilities. Pre-built experts handle domain-specific jobs like store review analysis, marketing copy generation, and operational data diagnostics for food delivery, service retail, and healthcare merchants. Teams can also record their own workflows to create custom skills.

Internally, Meituan already has 90,000 employees using 30,000 agents on the platform. The managed agent tier adds tenant isolation, credential management, hierarchical permissions, and a unified operations dashboard with full session traceability and cost monitoring.

Takeaways

— LongCat 2.0 is a 1.6T-parameter MoE model with 48B average active parameters and native 1M-token context, trained end-to-end on 50,000 domestic GPUs.

— CatPaw runs on mobile, desktop, and a 24/7 cloud mode that keeps long-running tasks alive even when local devices are offline.

— Agents can read files, control a browser, and execute terminal commands to produce deliverables like Excel reports, visualizations, or code.

— Multi-agent task decomposition lets several specialized agents work concurrently in isolated environments and auto-merge results.

— A marketplace of pre-built experts covers local-commerce tasks: store review diagnostics, marketing copy, campaign planning, and operational analytics.

— Teams can record browser workflows to auto-generate reusable skills, and the platform retains cross-session memory for user preferences.

— Managed Agents deploy as digital employees inside Feishu or WeCom, configurable via natural-language description without manual setup.

— Enterprise features include tenant-isolated runtimes, centralized credential management, hierarchical permissions, and private deployment support.

— Sandbox cold starts take hundreds of milliseconds, with dynamic scaling and idle release to manage cost.

— An operations dashboard provides full session logs, real-time debugging, and model-call cost tracking.

— Meituan reports 90,000 internal employees and 30,000 agents already running on CatPaw.

Conclusions

Training a 1.6T model entirely on domestic chips and then immediately productizing it as an agent platform closes the loop from silicon to application in a way most Western labs have not publicly demonstrated.

Bundling domain-specific experts for local commerce—rather than shipping a generic agent builder—signals that the moat for enterprise AI platforms is shifting from model size to vertical data and pre-integrated workflows.

The 30,000-agent internal deployment number is a meaningful signal of platform maturity; it suggests the orchestration, security, and observability layers have been stressed at scale before the public release.

Recording browser actions to generate reusable skills is a pragmatic alternative to purely prompt-based agent configuration and lowers the barrier for non-technical teams to encode institutional knowledge.

Concepts & terms

LongCat 2.0

Meituan's trillion-parameter Mixture-of-Experts language model with dynamic activation between 33B and 56B parameters, supporting 1-million-token context windows. It was trained entirely on domestic Chinese GPUs.

Managed Agents

CatPaw's enterprise tier that hosts AI agents as persistent digital employees inside collaboration tools like Feishu or WeCom, with tenant isolation, credential management, and an operations dashboard.

Sandbox environment

A lightweight, isolated runtime for each agent that cold-starts in hundreds of milliseconds and auto-releases when idle, balancing security and cost.

From the discussion

The visual identity invites an immediate comparison to WeChat. A more pointed thread questions whether the platform's real purpose is to optimize for the user or to extract more value from them, with one comment framing it as automated price discrimination and another wondering if it simply intensifies pressure on delivery workers.

— The platform's UI color scheme strongly resembles WeChat's.

— AI agents are framed as tireless labor that never clocks out, unlike human workers.

— A practical demand exists for an agent that handles the tedious work of price comparison and cart optimization for grocery orders.

— The platform is suspected of enabling algorithmic price discrimination against users.

— Concern is raised that the technology will be used to increase the workload and pressure on delivery riders rather than improve their conditions.

Featured comments

用户836060954012

Can you release an agent that recommends orders? I just wanted to buy a watermelon, but ended up spending ages searching for sellers, comparing prices, and building a cart to avoid the delivery fee.

八零后卷王

Smart price discrimination — smartly increasing the delivery riders' workload?

See top comments, translated →

Source: juejin.cn ↗ Google Translate ↗ Backup ↗