Meituan open-sources trillion-parameter LongCat-2.0, Citi says moat widens

Meituan open-sourced its trillion-parameter large language model LongCat-2.0 under an MIT license, giving away its most advanced AI to attract developers while betting its proprietary rider network data creates an unassailable moat in local services.

"LongCat-2.0's advanced Agentic architecture will enhance internal R&D efficiency and attract external developers, further fortifying Meituan's leadership in the local services market," Citi analysts wrote in a report, maintaining a Buy rating and HKD113 price target.

The model features a 1 million-token context window and is released with no restrictions under the MIT license, making it one of the most permissively licensed large models from a major Chinese tech company. Meituan shares rose as much as 5% following the announcement, according to CMSI, before settling 1.2% lower at the July 10 close with short selling accounting for 27.7% of turnover.

The open-source strategy pits Meituan against a crowded field of Chinese AI contenders — including ByteDance, Tencent with its Hy3 model, and MiniMax planning a 2.7 trillion-parameter M3 Pro — but Citi argues Meituan's advantage lies outside the model itself. Its rider network generates proprietary transaction and operational data that no general-purpose LLM provider can replicate, creating what the broker called a "competitive moat" that deepens relationships with merchants and consumers.

The trillion-parameter model joins a wave of Chinese AI releases that have accelerated since early 2026. Tencent's Hy3, which activates only 21 billion of its 295 billion total parameters, scored 75.8% on SWE-Bench Multilingual and 71.7% on Terminal Bench, matching larger rivals such as GLM-5.2 and GPT-5.5. MiniMax raised $2 billion in a funding round that was seven times oversubscribed, with CEO Yan Junjie pledging to forgo salary until AGI is achieved. Against these competitors, Meituan's edge is not raw benchmark scores but domain-specific data: its network of millions of delivery riders generates real-time information on traffic patterns, restaurant operations, and consumer behavior that can be fed into AI-driven marketing and business insights for small and medium enterprises.

For investors, the question is whether open-sourcing a trillion-parameter model creates measurable financial returns. Citi believes it will, through two channels: improved internal R&D efficiency that reduces costs, and increased merchant adoption of AI tools that boosts transaction volume and take rates. Meituan trades at roughly 18 times forward earnings, a discount to some Chinese tech peers, reflecting market uncertainty about margin pressure from food delivery competition and regulatory overhang. The open-source move could narrow that discount if it accelerates merchant monetization.

Risks remain. ByteDance has made inroads with its Seedance video generator and can provide strong AI support for local merchants through Douyin's ecosystem. The broader Chinese AI sector is fragmenting rapidly, with multiple trillion-parameter models competing for developer mindshare. Meituan's LongCat-2.0 must demonstrate that its agentic capabilities translate into real-world merchant outcomes — higher order volumes, lower customer acquisition costs — before the competitive moat Citi describes becomes visible in financial results.

This article is for informational purposes only and does not constitute investment advice.