Meituan

What it is

Meituan's LongCat team published the LongCat-Flash technical report on 1 September 2025^[1]. LongCat-Flash is a language model built on a mixture-of-experts (MoE) design with 560 billion parameters in total^[1].

In a mixture-of-experts model the network holds many separate expert sub-networks, and a routing step sends each token through only a small subset of them. The total parameter count therefore describes the model's full capacity, while the number of parameters actually used for a single token describes how much computation that token costs. For LongCat-Flash the second number is not fixed: the model activates between 18.6 billion and 31.3 billion parameters per token depending on context, averaging about 27 billion^[1]. That spread is the part of the design the report emphasises most.

How compute is allocated

The report describes the model as allocating its compute budget dynamically, and the mechanism for this is a set of zero-computation experts^[1]. These are experts the router can select but which perform no arithmetic, so the model can vary how much real computation a token receives without redesigning the routing machinery around it. In practice the same architecture can spend more computation on inputs that seem to need it and less on inputs that do not, and the average figure of about 27 billion parameters per token is the result of those choices across a workload.

The second design choice concerns how the model is spread across many devices. A shortcut-connected mixture-of-experts architecture enlarges the computation-communication overlap window, which raises inference throughput^[1]. In large models, a device frequently has to wait while other devices send it the data it needs; overlap means filling those waiting periods with useful arithmetic. Widening the window in which the two can be interleaved reduces idle time and is one route to faster serving. The report claims inference throughput above 100 tokens per second^[1].

Training

The team trained the model on more than 20 trillion tokens within 30 days^[1]. Getting a run of that size to finish in that window depends on the training stack, which the report describes as combining four things: hyperparameter transfer, model-growth initialisation, a stability suite, and deterministic computation for reproducible runs^[1].

The logic of each is straightforward. Hyperparameter transfer means settings such as learning rates are tuned on a smaller model and carried over to the large one, rather than being searched for at full scale. Model-growth initialisation means starting from a smaller trained model and expanding it instead of beginning from random weights. A stability suite is a set of checks and interventions that keep a long distributed run from diverging. Deterministic computation means the same inputs produce the same outputs, so a run can be repeated and a failure can be traced rather than guessed at.

Release

Weights are released under the MIT licence^[3], a permissive licence that allows reuse and modification with few conditions. The Hugging Face model card was created on 29 August 2025, three days before the report, and the repository points users to a chat endpoint at longcat.ai^[2].

Where it sits

LongCat-Flash belongs to a class of models that separate capacity from cost: the total parameter count is large, the per-token cost is a fraction of it, and the fraction can move with the input. What distinguishes this release within that class is the way the variable is implemented — a pool of zero-computation experts rather than a single fixed activation rate — together with a report that documents the training stack alongside the architecture. The MIT licence and the early model card make the weights reachable for outside developers at the same time as the technical description.