Hot on the heels of reports restricting internal Doubao use, Meituan open-sources its own model.

Meituan moved fast. In early July, the LongCat-2.0 LLM was officially open-sourced to considerable buzz in the open-source community. Just days earlier, reports emerged that Meituan had begun comprehensively restricting internal use of the Doubao LLM, requiring business units to migrate to LongCat, DeepSeek, and other models. From buying APIs to building models, from internal use to open-sourcing externally, Meituan's AI roadmap is becoming increasingly clear.

Why Is Every Big Tech Company Building Its Own Model?

Meituan isn't the first to walk this path. Alibaba has Tongyi Qwen, ByteDance has Doubao/Skylark, Tencent has Hunyuan, Baidu has ERNIE — almost every Chinese internet giant has a self-developed LLM. But previously the strategy was “self-develop + buy externally” walking on two legs: self-developed for important scenarios, third-party APIs for general use. Now the wind is clearly shifting: self-developed models are rising in priority, and third-party model usage is tightening.

The reasons come down to three. First is cost. When internal AI call volume reaches a certain scale, continuing to pay third-party model vendors per token delivers diminishing ROI. Doubao's 180 trillion daily tokens are primarily consumer traffic, but enterprise-level call costs are similarly staggering — Tencent giving developers 1,400 RMB/month starting token quotas is a clear cost signal. Training your own model and deploying it yourself yields much lower marginal costs long-term.

Second is data security. Alibaba's blanket Claude ban and Meituan's Doubao restrictions are both driven by core concerns about data leakage. Search and recommendation algorithms, merchant data, user behavior data, code — these are internet companies' core assets; inputting them into third-party LLMs carries real risk regardless of what confidentiality clauses say in contracts. Self-developed models deployed in your own server rooms, with data never leaving the intranet, solves this problem fundamentally.

Third is ecosystem leverage. According to Sullivan's latest full-stack AI cloud services report, Alibaba Cloud holds the #1 position with 40.1% market share. Models aren't just technology products; they're entry points for cloud services and ecosystems. Open-sourcing a model, attracting developers to use it, then funneling developers to your own cloud services — Alibaba Cloud has already proven this path works, and Meituan and Tencent are following.

Open Source Becomes Standard, but Monetization Remains Hard

LongCat-2.0's choice to open-source is now standard Big Tech model practice. DeepSeek built its name through open source; Tongyi Qwen's open-source versions enjoy strong community reputation; the Llama series proved open-source models' commercial value. Open sourcing itself doesn't generate revenue, but it builds developer ecosystems, collects feedback for model iteration, and lowers enterprise customer adoption barriers — essentially trading free models for market share and ecosystem position.

But monetizing open-source models remains an industry-wide challenge. Meta open-sources Llama without directly making money from the model, primarily for ecosystem and defensive reasons; DeepSeek is still exploring monetization paths; most open-source model projects make money through cloud services, enterprise editions, and technical support. Meituan's LongCat will face the same problem: open source can earn reputation, but how do you turn reputation into revenue?

For Meituan, though, LongCat first serves internal business — food delivery recommendations, in-restaurant dining, hotel and travel, instant retail. These scenarios generate massive daily AI call volumes; serving its own business well first is already a win through cost savings. Open-sourcing is icing on the cake, not the cake itself.

Starting in H2 2026, China's LLM market enters a new phase: the ‘audition’ phase for general-purpose LLMs is over, settling into a coexistence pattern of Big Tech self-built models + a handful of independent model companies (like DeepSeek, Moonshot, Zhipu). For startups hoping to build another general-purpose LLM to challenge Big Tech, the window of opportunity has essentially closed. The opportunities are in vertical scenarios, at the application layer, in Agent infrastructure — not in “building another GPT.”