| Field | Value |
|---|---|
| Type | AI model team inside a consumer-platform company |
| Parent | Meituan (HKEX-listed) |
| Country | China (Beijing) |
| Founded | March 2010 (Meituan); LongCat model line from late 2025 |
| Known for | LongCat open-weight model family; LongCat-2.0 (1.6T MoE, MIT license) |
Meituan is a Beijing-based local-services and delivery company that has become one of China's most active developers of open-weight large language models through its LongCat model family. Its LongCat-2.0, released June 30, 2026 under the MIT license, is a 1.6-trillion-parameter mixture-of-experts model that the company says is China's biggest model trained entirely on domestic chips (Source: scmp.com).
Company background
Meituan was founded in March 2010 by Wang Xing as a Groupon-style daily-deals site and, following a 2015 merger with review platform Dianping, grew into one of China's dominant "super apps" spanning food delivery, instant retail, hotel bookings, and local reviews. The company is publicly traded on the Hong Kong Stock Exchange and reports more than 770 million annual transacting users and over 14.5 million merchants (Source: venturebeat.com).
Facing domestic-market competition and margin compression in its core delivery business, Meituan publicly committed to investing "billions" in artificial intelligence and domestic chip capability, repositioning itself from a regional e-commerce and logistics operator toward AI infrastructure (Source: venturebeat.com).
The LongCat model family
Meituan's model line began in late 2025 with LongCat-Flash, a 560-billion-parameter mixture-of-experts foundation model, followed by the reasoning-focused LongCat-Flash-Thinking; both were released open-source under enterprise-friendly licenses (Source: venturebeat.com). Coverage of the open-model ecosystem describes Meituan as a constant releaser of new models and variants, including specialized lines such as LongCat-Flash-Omni (Source: interconnects.ai).
LongCat-2.0
LongCat-2.0, announced June 30, 2026 on GitHub, Hugging Face, and Meituan's LongCat platform, is the family's flagship: 1.6 trillion total parameters with roughly 48 billion active per token (dynamic activation between 33 and 56 billion depending on query complexity), a one-million-token context window, and an MIT license. At announcement, the company said full model weights were "coming soon" while API access was live (Source: venturebeat.com).
Before being claimed by Meituan, the model operated for roughly two months on OpenRouter as the anonymous stealth model "Owl Alpha," where it accounted for approximately 10.1 trillion monthly tokens (about 559 billion per day), a 242% month-over-month volume increase that placed it in the platform's global top three by usage (Source: venturebeat.com).
Meituan reports LongCat-2.0 scoring 59.5 on SWE-bench Pro (compared with 58.6 reported for GPT-5.5), 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual, and 73.2 on the FORTE workflow benchmark, while trailing premium closed models such as Claude Opus 4.8 on broader general-agent benchmarks (Source: venturebeat.com).
Training hardware and export-control context
Meituan says LongCat-2.0 was trained entirely on domestic Chinese chips — reported as a cluster of more than 50,000 domestic application-specific integrated circuits (ASICs) — rather than on Nvidia GPUs; the claim is the company's own, and initial coverage did not identify the specific chips (Source: scmp.com; venturebeat.com). If accurate, the release bears on the debate over whether US chip export controls constrain Chinese frontier training (see Export Controls (AI) and Semiconductor Supply Chain). The launch also coincided with US government restrictions on access to the newest American frontier models, which commentators argued was widening the opening for lower-cost Chinese open-weight alternatives (Source: venturebeat.com).
Commercial model
LongCat API pricing at launch was $0.75 per million input tokens and $2.95 per million output tokens (promotionally discounted to $0.30 and $1.20), with context-cache hits processed free of charge. Alongside pay-as-you-go billing, Meituan sells fixed-volume 30-day "Token Packs" released in limited flash sales four times daily (10:00, 16:00, 21:00, and 23:00 Beijing time), a structure the company presents as accommodating the burst compute patterns of autonomous coding agents (Source: venturebeat.com).
Relationships
- related: LongCat-2.0 (Meituan) — the company's flagship model
- related: Open-Weight Frontier Models, Fast-Follow Problem — LongCat-2.0 is a data point in the open-vs-closed capability-lag debate
- related: Export Controls (AI), Semiconductor Supply Chain — domestic-chip training claim
- related: Alibaba / Qwen Team, Zhipu AI, DeepSeek — peer Chinese open-weight model developers