--- title: "Meituan open-sources LongCat: a 560B model built to run agents, not small talk" date: 2026-10-15 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-meituan-longcat language: en --- # Meituan open-sources LongCat: a 560B model built to run agents, not small talk > The food-delivery giant released and open-sourced LongCat-Flash, a 560-billion-parameter mixture-of-experts large model (大模型) tuned for long-running agentic tasks — and shipped an API alongside it. Most people know Meituan as the app that brings lunch to the door in twenty minutes. Behind that mundane errand sits one of the hungriest AI labs in China, and this autumn it did something the model monopolies usually don't: it gave the weights away. The model is called LongCat-Flash. It is not a chatbot dressed up as a product — it is a large model (大模型) designed to run for a long time, call tools, and finish messy multi-step jobs that would time out a normal assistant. ## What actually shipped - LongCat-Flash-Chat, released and open-sourced on 1 September 2025, on GitHub and Hugging Face. - LongCat-Flash-Thinking, a "reasoning" sibling, released on 22 September 2025. - A paid API platform (LongCat API), live since 29 August 2025, plus a free chat surface at longcat.ai. So this was not a press release with a locked model behind it. Meituan published the weights, the code, and a way to pay for scale. ## The architecture bet: sparsity over size LongCat-Flash uses a mixture-of-experts (MoE) design. The headline numbers: - Total parameters: 560 billion. - Activated per token: between 18.6B and 31.3B, averaging about 27B. In plain terms: the model carries a huge brain but only wakes the parts it needs for each token. Meituan calls the dormant units "Zero-Computation Experts," and a controller keeps the average awake-set near 27B. The point is efficiency — you get near-frontier behavior while paying for a fraction of the machine. ## Why agents, specifically Meituan built its own evaluation set for agentic behavior and trained on synthetic trajectories generated by multiple agents. The stated goal is models that hold a task across many steps: open a browser, click, read, write, call an API, check the result, retry. That is a different bar from "write me a poem." A food-delivery platform lives on exactly those workflows — routing, scheduling, cancellations, refunds, merchant onboarding — so the model is shaped by the company's own operational scars. ## Speed as a feature Meituan claims two engineering results worth noting: - The model trained in about 30 days. - On H800 hardware it serves over 100 tokens per second per user. For a 560B-class model, both are aggressive. The later Flash-Thinking variant kept the speed while adding reasoning and, per Meituan, became the first domestic open model to pair deep thinking with tool use and both informal and formal (theorem-proving) reasoning. ## The context and language push By late September the Chat model doubled its context window from 128K to 256K tokens and added nine languages — Spanish, French, Arabic, Portuguese, Russian, Indonesian among them. That is a quiet signal: Meituan is not only optimizing for Chinese developers. ## How this fits Meituan's business Meituan is not OpenAI. It is a local-services company with millions of merchants and riders. An in-house large model (大模型) that is cheap to run and good at agentic loops is a direct cost lever: fewer humans on repetitive tickets, smarter dispatch, better merchant tooling. Open-sourcing also buys something less tangible — a developer ecosystem that stress-tests the model for free and builds the integrations Meituan would otherwise have to write alone. ## Why a delivery app, why now China's large model (大模型) race has split into two camps: the frontier labs chasing the biggest models, and the application companies that just need AI to work cheaply inside their own walls. Meituan sits firmly in the second camp, and LongCat is the proof that a non-lab company can ship a credible frontier-class model when the model serves its own operations. The open-source move also blunts a hiring problem. Top researchers want published, citable work; a locked internal model hides their contribution. Giving LongCat away is how Meituan recruits and retains without matching frontier-lab salaries. ## Honest limitations - Most benchmark claims here come from Meituan's own announcements and technical posts (meituan.com, longcat.ai). We did not independently reproduce the 100+ tokens/s or 30-day training figures. - "SOTA" and "rivals leading models" are Meituan's framing; independent third-party evaluations are still thin. - The model is tuned for agentic and Chinese-enterprise use. General creative or multilingual quality outside the nine named languages is unverified by us. - Open weights do not mean open training data; the underlying corpus is not public. - LongCat-Flash is a non-thinking base model by design — for hard math or code, the Thinking variant is the one to test. ## What readers can do now - Try it free at longcat.ai before trusting any vendor claim. - Pull the weights from Hugging Face or GitHub and run a small agentic task you already do by hand. - If you build for Chinese enterprise users, benchmark LongCat-Flash against Qwen and DeepSeek on your own tickets, not on leaderboards. - Watch the API pricing page; Meituan is pitching efficiency, so cost-per-task is the metric that matters. - If you only need chat, this is overkill — a smaller dense model will be cheaper and faster for you. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-meituan-longcat Free to quote with attribution and a link to the original.