Meituan launches LongCat Flash Chat with 560 billion parameters and a focus on fast AI agents

A large model built to respond quickly

Meituan has released LongCat Flash Chat, an open source artificial intelligence model with 560 billion parameters that uses only a small share of its network for each piece of text it processes. Announced in early September 2025, the model puts the Chinese on demand services company into a growing contest to build capable AI systems that respond quickly and cost less to run.

Meituan says the system typically activates between 18.6 billion and 31.3 billion parameters per token, averaging about 27 billion. Its technical notes cite generation speeds above 100 tokens per second for a single user on H800 class hardware, alongside an output cost near 5 yuan per million tokens under specific conditions.

The company positions LongCat Flash Chat as a production ready, nonthinking foundation model. A foundation model is a general AI system that developers can adapt to different applications. Here, nonthinking describes a focus on fast responses, instruction following and tool use, rather than the slower, extended reasoning associated with systems built to work through elaborate chains of logic.

Meituan's benchmarks put it alongside DeepSeek V3.1, Alibaba Cloud's Qwen3 family and Moonshot AI's Kimi K2, with competitive results against Anthropic's Claude Sonnet and Google's Gemini 2.5 Flash. Those comparisons come from the company's own testing.

The release follows Meituan's $281 million acquisition of AI startup Light Year in 2023. It also gives developers access to a core model from a company better known for food delivery and local services, extending its AI ambitions beyond features inside its own applications.

What developers can access

LongCat Flash Chat is available through GitHub and Hugging Face, with model weights, documentation and a technical report accompanying the release. Weights are the numerical settings a model learns during training. Publishing them lets other organisations run and adapt the system instead of relying entirely on a service controlled by its creator.

Developers can find the weights and documentation on the Hugging Face model page, while product information appears on the official LongCat website. The code and weights are released under the MIT License, which permits broad use, modification and commercial deployment.

The model supports a context window of up to 128k tokens. Tokens are the pieces of text an AI system processes, which may be words or parts of words. The context window is the amount it can consider at once, giving it room to work with lengthy documents, extended conversations and tasks involving several steps without constant restarts.

Caixin Global described the model as lacking advanced reasoning capabilities. Meituan's documentation nevertheless highlights competitive results across general benchmarks, instruction following, mathematics, coding, tool use and safety evaluations. Its intended strength is a combination of responsiveness and practical task handling.

How a network of experts saves work

LongCat Flash Chat uses a Mixture of Experts architecture. Instead of sending every token through the entire network, it divides parts of the model into specialised components called experts. A routing system chooses which experts to use for each token.

The idea resembles calling a few relevant specialists from a much larger team. The whole team supplies capacity, but only some members work on each item. Parameters are the learned numerical values within that network, so activating fewer parameters reduces the computation needed for each token.

Meituan's average of about 27 billion active parameters is roughly 4.8% of the model's 560 billion total. That comparison helps explain its design: a large network can hold substantial capacity while drawing on a much smaller portion during each processing step.

Meituan adds Zero Computation Experts, components that act as placeholders and can be selected when the router needs to balance workloads or reduce computation for simpler tokens. The design gives the model flexibility over how much work it performs.

A PID controller adjusts the bias used in expert selection to keep the average number of active parameters steady. PID stands for proportional, integral, derivative. It is a control method that nudges a changing system toward a target while limiting repeated swings above and below it.

For LongCat Flash Chat, that target concerns how much of the network is active. Together, the router, placeholders and controller are intended to help the model distribute work while keeping its computing demands predictable.

Where the speed and cost claims come from

Meituan also uses a shortcut connected MoE layout to increase the overlap between communication and computation. Large models often run across several graphics processing units, or GPUs. Those processors must exchange information, and time spent moving data can slow the whole system.

The shortcut connections create paths that allow parts of the network to progress while information travels between devices. The aim is to keep processors working instead of waiting for one another. The technical notes cite more than 100 tokens per second for a single user on H800 class hardware.

Open source summaries and Chinese media coverage say training took about 30 days using a highly parallel setup, meaning many computing resources worked at the same time. They also describe generation at 100 tokens per second on H800 accelerators.

Meituan's output cost claim is near 5 yuan per million tokens under specific conditions. Actual deployment costs vary with the hardware, the number of requests processed together and the software used to serve the model. The release presents speed and cost as linked benefits of activating only part of the network for each token.

How Meituan trained for reliability

Meituan's repository describes several methods intended to make training stable. Hyperparameter transfer reuses useful training settings across model sizes. Model growth initialisation lets larger versions begin from smaller ones, providing a starting point for further training.

The team also uses a router gradient balancing method with z loss to support the routing system. In plain terms, these are training controls intended to keep the mechanism that assigns work to experts operating reliably as the model learns.

Deterministic computation is another stated priority. It supports reproducible runs, allowing the team to repeat computations consistently when checking results and investigating problems.

For training data, Meituan describes a two stage pretraining fusion strategy and a framework that uses multiple agents to synthesise complex task trajectories. A task trajectory is a sequence of actions used to complete a job. Training on such sequences supports workflows in which a model must choose tools and carry out several connected actions.

These choices fit the product's stated emphasis on reliable agent behaviour. Meituan is directing the training pipeline toward practical task execution and responsive service, with less emphasis on deep, extended reasoning.

How it compares with rival models

Meituan's technical report places LongCat Flash Chat on a similar performance level to DeepSeek V3.1, Alibaba Cloud's Qwen3 family and Moonshot AI's Kimi K2. It also describes strong results against Anthropic's Claude Sonnet and Google's Gemini 2.5 Flash. Some coverage includes comparisons with GPT 4.1 for particular nonthinking tasks.

The exact ranking depends on the benchmark suite and prompt settings. A benchmark is a set of tasks used to measure a model's performance, while prompt settings influence how a question or instruction reaches the system. Different tests can reward different abilities.

Meituan's results are company assessments; independent evaluations would give a fuller picture of performance across applications. The release does not supply one universal ranking covering every kind of AI work.

Reasoning oriented systems devote more processing to planning and reflecting across multiple steps, often taking longer and using more computation. LongCat Flash Chat instead prioritises rapid instruction following, tool use and the coordination of agent workflows.

For businesses, that choice makes response time, throughput and predictable costs central measures of usefulness. Throughput describes how much work a system can complete in a given period. A service handling many routine requests may place particular value on a model that responds promptly and keeps requests moving.

What agents could do in everyday services

An AI agent uses a model to choose actions, call tools and follow a workflow. It might retrieve information, consult an internal database, browse documentation, trigger a payment API or draft a customer message. An API is a connection that allows one piece of software to request an action or information from another.

LongCat Flash Chat is tuned for this type of work, combining tool use with a long context window and low latency. Latency is the delay before a system responds. Keeping that delay short matters when an application must carry out several actions before returning an answer.

Meituan has already showcased NoCode for AI assisted coding, Kangaroo Advisor for business decision support and Meituan Jibai for hotel operations management. Those products show the range of business activities around its broader agent efforts.

A fast agent engine could support merchant onboarding, customer support triage, inventory checks, route planning assistance and marketing copy generation. These are potential applications described around the model's capabilities; the release does not specify which of these workflows already use LongCat Flash Chat.

In local services and electronic commerce, many jobs involve finding information, choosing an action and passing results between systems. Meituan's model is aimed at that practical combination of speed, tool connectivity and the ability to retain a long working context.

A foundation model strategy takes shape

Meituan's purchase of Light Year in 2023 for $281 million marked an early step in its push into core AI models. Company leaders have described three layers of activity: using AI for internal work, putting AI into consumer products and building foundational models.

LongCat Flash Chat is the first public proof point described at that foundational layer. The sequence connects an acquisition and a broader company strategy with a model that outside developers can now examine and deploy.

• 2023: Meituan acquired AI startup Light Year for $281 million.
• Early September 2025: Meituan announced LongCat Flash Chat and made the release available through GitHub and Hugging Face.
• September 5, 2025: The launch includes model weights, documentation, a technical report and deployment guidance.

The release also joins a wider wave of open source models from Chinese companies, including Alibaba, labs affiliated with DeepSeek and Moonshot AI. Rapid iteration and community testing are helping those developers narrow performance gaps with Western competitors.

Meituan's permissive licensing invites startups, researchers and enterprise teams to experiment with the model and adapt it for their own products. That can bring feedback and new applications while building technical credibility beyond the company's food delivery business.

Coverage of the strategy describes Meituan's approach as aggressive and proactive, with speed and openness central to attracting developers. Publishing a capable model gives those developers another potential supplier for applications that need quick responses and agent capabilities.

Deployment, licensing and sensitive uses

The MIT License allows broad commercial use and modification of LongCat Flash Chat. The repository includes deployment guidance for vLLM and SGLang, frameworks used to run language models efficiently and handle large volumes of requests.

Those frameworks offer techniques such as continuous batching and paged attention. Continuous batching helps keep processors occupied by managing requests together as work arrives and finishes. Paged attention manages the memory used while processing context. Both can help reduce the cost of serving a model.

A model that activates only some experts for each token fits an operating environment designed to reduce idle time on GPUs. Its actual performance will still depend on the deployment framework and hardware configuration.

Meituan's documentation urges developers to assess safety, fairness and legal compliance in their own settings. The release includes safety evaluations and red teaming, a process of testing a system with challenging or adversarial requests to identify weaknesses.

Applications involving personal information, finance or health require particular care. Meituan's guidance calls on users to follow applicable laws and industry rules when putting the model into sensitive settings.

Investors look beyond the launch

Meituan's shares traded lower after the release, according to financial coverage. TipRanks cited benchmark comparisons discussed by the South China Morning Post and placed the announcement against a wider backdrop of Chinese support for the technology sector and continuing trade frictions.

The financial coverage gave no size for the share price decline. Investor attention now centres on adoption, savings within Meituan's operations and whether outside developers choose the model for working products. Market reactions can lag the practical results of a product launch.

For enterprise customers, the launch adds a model supplier from outside the group of companies primarily identified with AI development. Its pitch is concrete: a large network, selective computation, quick responses and support for agents that use tools.

Meituan's commercial task is to turn those technical claims into useful services. The wider competitive pressure is also practical, pushing model developers to show value through response times, operating costs and completed workflows alongside benchmark scores.

Key Points

• Meituan released LongCat Flash Chat in early September 2025 with 560 billion total parameters.
• The model typically activates 18.6 billion to 31.3 billion parameters per token, averaging about 27 billion, or roughly 4.8% of its total.
• It supports a context window of up to 128k tokens and prioritises fast instruction following, tool use and agent workflows.
• Meituan cites speeds above 100 tokens per second on H800 class hardware and an output cost near 5 yuan per million tokens under specific conditions.
• Company benchmarks compare it with DeepSeek V3.1, Qwen3, Kimi K2, Claude Sonnet and Gemini 2.5 Flash.
• The code and weights use the MIT License, with deployment guidance for vLLM and SGLang.
• The launch follows Meituan's $281 million purchase of Light Year in 2023 and supports its strategy across internal work, consumer products and core models.
• Adoption, operating savings and use by outside developers are the commercial tests following the release.