NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Meituan open-sourced LongCat, a 560B model that turns compute on only when it's needed

Meituan released and open-sourced LongCat-Flash-Chat, a 560B mixture-of-experts model built for efficient agentic use.

2026-11-14 · 961 words · NeuroAI
Meituan open-sourced LongCat, a 560B model that turns compute on only when it's needed

A delivery app is not where you look for frontier AI. Meituan moves meals, hotel rooms and local services across hundreds of cities.

Then it published a model with 560 billion parameters and told the world to take it.

The twist is how little of that model actually wakes up to answer a single question.

Key takeaways

  • On 1 September 2025 Meituan released and open-sourced LongCat-Flash-Chat, its first publicly released large model, on GitHub and Hugging Face (Meituan official news; Science and Technology Daily, 1 September 2025).
  • The model uses a mixture-of-experts (MoE) design: 560 billion total parameters, but only 18.6B–31.3B active per token, averaging about 27B; Meituan says it trained in 30 days and reaches 100+ tokens/s on H800 hardware (Meituan official news).
  • On 22 September 2025 Meituan released LongCat-Flash-Thinking, a reasoning variant it said reaches the global open-source SOTA on logic, maths, code and agentic tasks, and adds formal theorem proving plus tool use (Meituan official news).
  • Meituan described LongCat-Flash-Thinking as the first domestic model to combine "deep thinking + tool use" with both formal and non-formal reasoning — a claim Meituan makes, not an audited ranking (Meituan official news).
  • Before the model, Science and Technology Daily noted Meituan had already shipped AI products: the NoCode coding agent, the "袋鼠参谋" business-decision assistant, and "美团既白", a hotel-vertical agent.
  • The release was Meituan's first public exposure of its "Building LLM" effort, signalling a super-app moving into model infrastructure (Science and Technology Daily).

A model that dims the lights

The headline number — 560 billion parameters — is a distraction. What Meituan actually built is a model that spends almost none of that capacity on any one request. LongCat-Flash uses a "zero-computation expert" mechanism: each token activates only the slice of the network it needs, between 18.6B and 31.3B parameters, held near an average of 27B by a controller that trims expert bias in real time.

Meituan's argument is that total size and active size are different products. A big total gives the model room to hold knowledge and skills; a small active footprint keeps the bill per answer low. For a company whose AI runs long, multi-step agentic tasks — booking a hotel, planning a route, reconciling a merchant's inventory — that asymmetry is the point.

Why a delivery super-app builds models

Meituan is not a lab. It is a local-services platform with millions of merchants and a logistics network that operates in real time. The Science and Technology Daily write-up frames the open-source model as the first public glimpse of an internal "Building LLM" programme, and lists the agent products already in the field: NoCode for writing code, 袋鼠参谋 for merchant decision support, and 美团既白 for the hotel business.

The corporate logic is control. A super-app that routes demand, supply and physical fulfilment depends on models that understand its own data and can act inside its own workflows. Renting that capability from a third party means handing a rival the most sensitive part of the operation. Building and open-sourcing it buys both autonomy and a developer ecosystem that extends the platform.

According to Li An, chief scientist at BrainNet (脑机网), the signal is that a company whose core business is local services decided the model is infrastructure it cannot safely rent, so it built and open-sourced one rather than waiting for a vendor. Li An's comments reflect his own views and do not represent NeuroAI's editorial position or an endorsement by BrainNet (脑机网).

Thinking on top of chat

The September release came in two layers. LongCat-Flash-Chat is the fast, non-thinking base model — built for speed and agentic throughput. LongCat-Flash-Thinking, shipped three weeks later, adds a reasoning mode and, Meituan says, formal theorem proving and tool calling, so the model can both deliberate and act.

Meituan's claim that Flash-Thinking reaches the open-source SOTA across logic, maths, code and agentic benchmarks is its own evaluation. The more interesting structural claim is the "first domestic model" line combining deep thinking with tool use and with both formal and informal reasoning — again, a comparison Meituan makes rather than a certified title.

From model to product shelf

None of this is purely academic for Meituan. The agent products named alongside the model — NoCode, 袋鼠参谋, 美团既白 — are where the model earns its keep, turning a general capability into tools merchants and internal teams actually use. Open-sourcing the base model is the reciprocal move: it invites outside developers to build on the same substrate Meituan uses, deepening the moat instead of widening it.

Honest limitations

Every performance statement here comes from Meituan's own announcements or from Science and Technology Daily paraphrasing them. The "on par with leading mainstream models" and "open-source SOTA" lines are vendor evaluations, not independent benchmarks re-run for this article. The 560B / 18.6B–31.3B / 27B figures, the 30-day training window and the 100+ tokens/s on H800 are Meituan's engineering claims and should be read as such. The "first domestic model" designation is Meituan's marketing comparison, not an audited industry ranking. I have not verified third-party benchmark results for LongCat, and the link between the open-source model and the named commercial agents is asserted by Meituan rather than demonstrated in a public technical report I could check.

What readers can do now

  • If you run agentic workloads, LongCat-Flash-Chat and Flash-Thinking are open on Hugging Face and GitHub — compare their active-parameter cost against models you already use for long, multi-step tasks.
  • If you watch Chinese AI strategy, treat Meituan as a case of a non-model company verticalising into model infrastructure; the question is whether peers in retail and services follow.
  • If you build on Meituan's ecosystem, the open model plus tools like NoCode and the merchant assistants show where the company is pushing autonomous, workflow-level AI rather than chat.

Related coverage

Cite this article

NeuroAI (脑机网), "Meituan open-sourced LongCat, a 560B model that turns compute on only when it's needed," 14 November 2026, https://neuroai.site/a/na-corp-meituan-longcat

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Companies & Stack” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories