LongCat-Flash

What it is

LongCat-Flash is a large language model built on a Mixture-of-Experts (MoE) design, with 560 billion parameters in total ^[1]. In an MoE model the heavy feed-forward layers are split into many separate "expert" sub-networks, and a router sends each token to only a few of them, so the total parameter count can be very large while the computation spent on any single token stays comparatively small. In LongCat-Flash that per-token budget is not fixed: depending on the context, the model activates between 18.6 billion and 31.3 billion parameters, averaging about 27 billion ^[1].

Where it comes from

Meituan's LongCat team published the LongCat-Flash technical report on 1 September 2025 ^[1]. The model card for the chat version was created on Hugging Face on 29 August 2025 ^[3], and the weights are released under the MIT license ^[2]. The accompanying repository points users to a chat endpoint at longcat.ai ^[2].

How it works

The most distinctive part of the design is a set of zero-computation experts ^[1]. Ordinary experts each perform a learned transformation; these experts perform none, so routing a token through them is a way of spending less compute on it. Because the router can mix real experts with zero-computation ones, the model can allocate its compute budget dynamically rather than spending the same amount on every token ^[1].

The architecture is also shortcut-connected: skip-style connections across the MoE layers enlarge the window in which computation and communication between devices can overlap, which raises inference throughput ^[1]. This matters because in a distributed MoE system the cost of moving activations between devices can dominate, and overlapping it with local computation is one of the main ways to keep generation fast.

Training such a model required a stack built for reproducibility as much as for speed. The team combined hyperparameter transfer, model-growth initialization, a stability suite and deterministic computation so that runs can be reproduced ^[1]. Hyperparameter transfer means settings tuned at small scale are carried over to the large model instead of being searched again from scratch; model-growth initialization means training can start from a smaller model and grow; a stability suite is a set of checks and interventions that keep a long training run from diverging; deterministic computation means the same inputs produce the same outputs, making failures diagnosable.

Training and reported results

The team trained the model on more than 20 trillion tokens within 30 days ^[1]. The report claims inference throughput above 100 tokens per second ^[1]. Alongside computational efficiency, the team frames the model around agentic capabilities — that is, use in settings where the model is expected to carry out multi-step tasks rather than only answer single questions ^[1].

Where it sits

LongCat-Flash belongs to a generation of open-weight MoE models in which the headline parameter count is much larger than the number of parameters actually used per token. That gap is the point of the design: it lets a model hold a great deal of stored knowledge while keeping the cost of generating each token manageable. The dynamic range of 18.6 to 31.3 billion active parameters, together with zero-computation experts, is an explicit attempt to make that trade-off adjustable at inference time rather than fixed in advance ^[1].

Releasing the weights under the MIT license, together with a hosted chat endpoint, gives other researchers and developers two ways to use the model: running it themselves or calling it as a service ^[2]. The technical report documents the architecture and training stack so that the choices behind the model can be examined and reused ^[1].