$ cat wiki/entities/meituan.md
Meituan
entityupdated 2026-07-05created 2026-07-05
Latest
- 2026-06-30
LongCat-2.0 open-sourced
Overview
Chinese consumer technology conglomerate (food delivery, local services, travel). In 2026, Meituan's internal AI research team ("LongCat AI") released LongCat-2.0 — a 1.6-trillion-parameter open-source model trained entirely on domestic Chinese chips, which became the largest publicly open-source MoE model at release.
Meituan is one of China's largest tech companies (comparable to Uber Eats + Yelp + Expedia combined), and its AI research capability — previously focused on recommendation systems and logistics optimization — has now produced a frontier-scale language model.
Key People
- LongCat AI team — internal AI research unit responsible for the LongCat model series (leadership names not publicly disclosed)
Models & Products
- LongCat-2.0 — 2026-06-30, 1.6T MoE, MIT license, trained entirely on Huawei Ascend 910, 1M ctx, agentic coding; leading OpenRouter as "Owl Alpha" before reveal
Recent Activity
- 2026-06-30: LongCat-2.0 open-sourced — 1.6T MoE, MIT license, trained on 50,000 Huawei Ascend 910 chips (no NVIDIA). The model had been quietly topping OpenRouter under codename "Owl Alpha." A 1M-token context window enabled by LongCat Sparse Attention (LSA). Why it matters: first demonstration that a 1.6T-scale training run can complete on non-NVIDIA domestic Chinese hardware, directly contradicting narratives that US chip export controls prevent frontier AI development inside China. → LongCat-2.0 (source) (VentureBeat)
Strategic Position
- Unlike dedicated AI labs (Baidu, Alibaba/Qwen, ByteDance/Doubao), Meituan's AI capability emerged from its core operational business (recommendation, logistics, pricing)
- The open-source MIT release under a Western-friendly license is notable — it maximizes adoption beyond China and contrasts with models like GLM-5.2 (Tsinghua-based Z.ai), which have more restrictive licensing
- The Ascend 910 training achievement has policy relevance: it is evidence that China's domestic chip ecosystem can support 1.6T-parameter training runs without NVIDIA hardware
Related
- LongCat-2.0 — LongCat-2.0 (1.6T MoE, June 30, 2026)
- Alibaba / Qwen AI Lab — Qwen series (comparable Chinese open-source competitor)
Open Questions
- Is "LongCat AI" an independent entity or a wholly internal Meituan team?
- Will Meituan release LongCat-2.0 as a commercial API?
- What are LongCat-2.0's verified benchmark scores vs. other frontier models?