Most big models are built to top leaderboards, then quietly cost a fortune to keep running. Meituan — a name most people associate with food delivery, not frontier research — just went the other way. It open-sourced a model designed to keep working cheaply inside real products.
The release is LongCat-Flash-Chat (龙猫 / LongCat), and the interesting part is not its size but how little of it runs at any moment.
What LongCat actually is
LongCat-Flash is a large model (大模型) built on a Mixture-of-Experts (MoE) architecture. The headline number is a total of 560 billion parameters, but the model only wakes up the parts it needs. Per token, it activates between 18.6 billion and 31.3 billion parameters, averaging around 27 billion.
That design choice matters more than the raw count:
- Zero-Computation Experts — a mechanism that lets the model skip entire expert blocks when a token does not need them, paying compute only for what is used.
- Cross-layer channels — MoE communication and computation run partly in parallel, which is what makes both training and inference faster.
- PID-controlled load balancing — during training, a controller nudges expert bias in real time to keep the average activation near 27B and avoid a few experts doing all the work.
Meituan says the model was trained in about 30 days and reaches more than 100 tokens per second for a single user on Nvidia H800 hardware. For a model this large, that inference speed is the real product story.
Why "agentic" is the point
LongCat-Flash is described by Meituan as a non-thinking base model — meaning it is not built around long chain-of-thought reasoning, but around doing things. The team built its own agentic evaluation sets and generated diverse, high-quality trajectory data using multi-agent methods.
The pitch is that long, multi-step agent workflows — booking, browsing, calling tools, filling forms — are where the cost of a full-size model quietly adds up. A model that activates ~27B of 560B per token is positioned to survive those long sessions without melting the budget.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the more interesting trend is not another giant model but Chinese internet-platform companies turning their inference-cost problems into open, reusable architectures.
Where it is available
LongCat-Flash-Chat is open-sourced on GitHub and Hugging Face, with a chat front-end at longcat.ai. Meituan also ships a fine-tuning toolkit and industry templates aimed at local-life services, customer support, and similar verticals.
This fits Meituan's stated "Building LLM" pillar — the company has also shipped an AI coding agent (NoCode) and an operations assistant (袋鼠参谋 / "Kangaroo Staff Officer"), so the open model is arguably the foundation layer under those products.
For a delivery-and-local-services company, an open model is also an ecosystem play: external developers who build on LongCat today may become partners or future hires. That logic — open the core, grow the periphery — is increasingly how Chinese platform companies approach foundation models, and it explains why a non-research brand would bother publishing weights at all.
How it sits in China's model landscape
China's open-weight ecosystem has moved fast, and LongCat joins a crowded field of domestically released models. What distinguishes it is the explicit cost-per-token framing. Meituan cites an output cost of about RMB 5 per million tokens (≈ US$0.70 / HK$5.4), a number aimed squarely at teams running agents at scale rather than hobbyists testing prompts.
These foundation models are part of a broader Chinese push that also spans humanoid (人形机器人) robotics and brain-computer interface (脑机接口) research, but LongCat keeps its feet firmly in software: cheaper, faster inference for task-completing agents.
What to watch as the open model matures
The figure worth tracking is not parameter count but total cost of ownership for a running agent. A 560B model that activates ~27B per token only pays off if teams actually deploy long, tool-calling sessions where per-token fees compound; for short prompts, a smaller dense model may still win. The honest test is whether third-party benchmarks confirm Meituan's speed and cost claims on real workloads, and whether the fine-tuning toolkit and industry templates lower the barrier enough for local-life and support teams to adopt it without a research staff. If the ecosystem grows, LongCat's significance will be as a reusable inference-cheap layer, not as another leaderboard entry.
Honest limitations
- Not a reasoning model. LongCat-Flash is a non-thinking base model; for hard math or code you should not expect it to replace a thinking/reasoning-tier model. Meituan later released LongCat-Flash-Thinking for that tier.
- Claims are vendor-reported. The 100+ tps and 30-day training figures come from Meituan's own technical write-up; independent benchmarks are still maturing.
- Scale still needs serious hardware. "Lean per token" does not mean "runs on a laptop." This is a 560B model requiring serious GPU infrastructure to serve at the quoted speed.
- Ecosystem is young. Tooling, community examples, and third-party benchmarks are far smaller than for longer-established open models.
What readers can do now
- Pull the weights and run a small agent. Clone the Hugging Face repo and test a tool-calling loop (web search, form fill) to see if the cost/latency profile holds for your workload.
- Benchmark your own agentic tasks. Don't trust vendor numbers — measure tokens-per-dollar on a real multi-step workflow before committing.
- Watch the Thinking variant. If your use case needs reasoning, track LongCat-Flash-Thinking rather than the base Chat model for head-to-head comparisons.