--- title: "Meituan's LongCat goes open-source: a 560B-parameter bet from a food-delivery app" date: 2026-10-21 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-meituan-longcat-open-source language: en --- # Meituan's LongCat goes open-source: a 560B-parameter bet from a food-delivery app > The company behind 30-minute meal delivery has released LongCat (龙猫), an open-weight MoE model, and is folding AI into merchants, riders and users. A delivery rider in Shenzhen taps a voice assistant at 2 a.m. to catch a restaurant's last booking. Across the city, a blind massage-shop owner finally reads his daily sales chart through a screen reader that makes sense of the numbers. Neither of them knows the same company just gave away a 560-billion-parameter AI model for free. That company is Meituan (美团), best known outside China for restaurant delivery and local services. In September 2025 it published and open-sourced LongCat-Flash-Chat (龙猫), and a few weeks later LongCat-Flash-Thinking. In July 2026 it said it had open-sourced LongCat-2.0. A firm whose core business is getting food to doors has quietly become a foundation-model developer. ## Why a delivery company built a foundation model Meituan is not a lab spun out of a university. It is a "retail + technology" (零售 + 科技) platform with hundreds of millions of local-life users, millions of merchants and a fleet of riders and autonomous vehicles. The case for building its own model is practical: the queries, menus, maps, delivery routes and merchant pain points it sees every day are a training ground that few pure AI startups can match. The company has described its AI push as resting on three layers — AI at work (internal efficiency), AI in products (features users and merchants touch), and Building LLM (the underlying model itself). LongCat is the third layer made visible. ## What LongCat (龙猫) actually is LongCat-Flash-Chat is a Mixture-of-Experts (MoE, 混合专家模型) model. The headline numbers, taken from Meituan's own technical release: - **Total parameters: 560 billion**, but only a slice activates per token. - **Activated parameters: 18.6B–31.3B per token, averaging about 27B** — the model spends compute only where the context needs it. - A mechanism Meituan calls **zero-computation experts (零计算专家)**, plus a PID controller that keeps average activation near 27B during training. - Trained in roughly **30 days**, and built to run fast: Meituan reports **100+ tokens per second per user on H800** hardware. - Released on GitHub and Hugging Face, with a chat site at longcat.ai. A non-reasoning base model first, then LongCat-Flash-Thinking (released September 22, 2025) added a "thinking" mode. Meituan says it reaches state-of-the-art among open models on several reasoning, math and code benchmarks, and claims it is the first domestic model combining deep reasoning with tool use and both informal and formal logic. The latest step came in July 2026, when Meituan said it had open-sourced **LongCat-2.0**: a 1.6-trillion-parameter model it describes as trained and served on a 50,000-card domestic compute cluster, released with weights, inference engine and documentation under an MIT license for free commercial use. Whether that scale holds up under outside testing is a separate question (see limitations), but the direction is clear — this is not a one-off demo. ## The merchant, the rider, the user The model is the substrate; the products are where it shows up: - **袋鼠参谋 (Kangaroo Staff)** — an AI decision assistant for restaurant merchants, covering category analysis, site selection and dish development. Meituan says it has been offered free to merchants and used to generate hundreds of thousands of reports. - **智能掌柜 (Smart Manager)** — a dine-in operations tool rolled out to cooperate restaurants, handling set-menu design, front-desk reception and review replies. - **美团既白** — a hotel-vertical AI agent for dynamic pricing and operations. - **小美 (Xiaomei)** and **问小团 (Ask Xiaotuan)** — consumer-facing AI search and task helpers inside the Meituan app. - **NoCode** — an AI coding agent the company shipped in 2025. - Voice and accessibility tools that answer late-night merchant calls and read dashboards aloud for visually impaired shop owners. The pattern is consistent: the model is trained once and then pushed into the three surfaces Meituan already owns — the merchant back office, the rider's handheld, and the customer's app. ## The strategic logic Meituan's move resembles what Alibaba did with Qwen and what Baidu did by folding Ernie into search: a company with a massive distribution advantage stops renting intelligence and starts owning it. The bet is that local-life services — messy, geographic, inventory-heavy, language-dense — are exactly the kind of problem where a general model trained elsewhere underperforms a model tuned on your own data. Open-sourcing is the lever. By releasing weights, Meituan gets external stress-testing, developer goodwill and a recruiting signal, while keeping the highest-value application layer (merchant tools, logistics, consumer features) for itself. ## Honest limitations - **Benchmark claims are the company's own.** The 560B/27B figures and the "SOTA among open models" language come from Meituan's releases. Independent, reproducible evaluations of LongCat-2.0's trillion-scale claims are still thin. - **Open weights ≠ open training.** Meituan has published model weights and code, not the full training data, recipe or the domestic-cluster details behind LongCat-2.0. "Fully open" here means usable, not fully reproducible. - **No disclosed AI R&D budget.** Unlike listed chip or cloud firms, Meituan has not broken out how much it spends on foundation models, so the cost discipline of this strategy is unverified. - **Revenue impact is indirect.** LongCat shows up as better merchant and user tools, not a separate line item, so its payoff is hard to isolate from the rest of the business. ## What readers can do now - **Try it directly.** Pull LongCat-Flash weights from Hugging Face or GitHub, or chat at longcat.ai, and compare agentic tasks (multi-step tool use) against Qwen or DeepSeek on the same prompt. - **If you run a local-service business, test the free tools.** 袋鼠参谋 and the voice receptionist are free to merchants and worth a controlled A/B against your current workflow before paying for anything. - **Watch the July 2026 milestone.** If LongCat-2.0's 50,000-card domestic-cluster claim is independently reproduced by developers, it signals a real shift toward large models trained outside US hardware — track community benchmarks, not press lines. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-meituan-longcat-open-source Free to quote with attribution and a link to the original.