Meituan releases LongCat-Flash

Two dates, one release

On 1 September 2025, the LongCat team at Meituan released LongCat-Flash together with a technical report describing the model ^[1]. The release also shipped open weights under the MIT license, one of the more permissive open-source licenses ^[2].

A detail in the timeline is worth noting: the Hugging Face model card for LongCat-Flash was created on 29 August 2025, a few days before the technical report appeared ^[2]. From the outside, that ordering suggests the weights were treated as the main artifact of the release, with the documentation following them rather than the other way round.

What the parameter numbers describe

The model has 560 billion total parameters, and activates between 18.6 billion and 31.3 billion of them per token ^[1]. The two figures measure different things. Total parameters are the full store of learned weights that has to be held in memory; activated parameters are the part that actually takes part in the arithmetic for a given token, and they set most of the compute cost.

When the activated figure is far below the total, the model is sparse: it keeps a large pool of parameters and routes each token through a fraction of them. The benefit is that a large parameter count can be supported without paying for all of it on every token. The cost is engineering complexity — routing decisions, load balancing and memory management all become harder problems. The fact that the activated figure is a range rather than a single number says the amount of computation spent on a token is not fixed; some tokens are routed through more parameters, others through fewer.

Training

Meituan said the model was trained on more than 20 trillion tokens and that the run completed within 30 days ^[1]. A token is the unit these models read and predict, roughly a word fragment, so 20 trillion tokens is an enormous amount of text and code. Thirty days for a run of that size implies a tightly engineered training stack. Both figures are the team's own account of its training run, not an outside measurement.

Throughput

The team reported throughput above 100 tokens per second ^[1]. Throughput matters differently depending on how a model is used. In a chat interface it sets how quickly text appears on screen. In an agent setting, where a model may be called many times in a row to finish one task, the latency of each call compounds, and generation speed becomes part of whether the whole task finishes in an acceptable time.

Positioning: built for agents

Meituan positioned LongCat-Flash for agentic use cases, not only chat ^[1]. The distinction is about what the model is expected to do. A chat model answers a question in one pass. An agent-style system is expected to work through multi-step tasks, call external tools, and use what those tools return to decide what to do next. Positioning a model this way is a statement about intended use, not a benchmark result.

Why the release matters

What makes the release notable is the combination of conditions. A model with 560 billion total parameters is expensive to train and heavy to host. Publishing the weights under the MIT license lets other organizations download, modify and redistribute them, and run them on their own hardware ^[2]. Open weights also make a model inspectable in a way a hosted API does not: outside researchers can run it themselves rather than reason about it from a description.

For a large model to arrive with weights, a permissive license, a stated target of agentic work and an unusual efficiency claim about training and serving is a rare package. The details above are the ones the release itself put forward; how the model performs in independent hands is a separate question, and one that open weights make it possible to ask.