Meituan Launches LongCat 2.5, a Huge Open Model for Coding and Visual Tasks

Meituan has released a new large artificial intelligence model called LongCat-2.5, and the numbers behind it are striking. The model was released on November 2, 2026. It has 1.6 trillion total parameters, with about 48 billion active for each token it processes. In simple terms, it is a very large system that only uses a slice of its full capacity at any one moment, which helps keep it fast enough to use in practice.

Meituan is best known to many readers as a major Chinese technology and services company, with food delivery and local services at its core. Its move into large open AI models shows how far that business has widened. LongCat-2.5 is aimed at two jobs that developers care about most right now: extended coding work and visual understanding. That means writing and fixing code over long projects, and making sense of images, charts, documents and other visual material.

The release also stands out for its scale of context. LongCat-2.5 supports a 1 million token context window, with a maximum output of 128,000 tokens. A token is roughly a small chunk of text, so a 1 million token window can hold a very large codebase, a long report or many documents at once. For developers, that could mean fewer workarounds when asking the model to work across a whole project instead of one file at a time.

Quick Facts

What 1.6 Trillion Parameters Actually Means

Parameters are the adjustable values inside an AI model that it learns during training. More parameters usually mean a model can capture more patterns, but they also mean higher training and running costs. LongCat-2.5 uses a design where only about 48 billion parameters are active for each token, even though the total is 1.6 trillion. This is often called a mixture of experts style approach, where different parts of the model specialise in different kinds of work and only the relevant parts switch on for a given task.

For users, the practical point is balance. A model can be huge in total size yet still respond at a reasonable speed and cost, because it does not use every part of itself for every word. That balance is one reason very large models have become more common in the past two years. The challenge is making sure the right experts are chosen for each task, so quality stays high across coding, reasoning and visual work.

Why a 1 Million Token Context Window Matters for Coding

Coding is rarely about one short snippet. A real project has many files, libraries, tests and notes, and a change in one place can break something in another. Older models forced developers to paste in small pieces at a time, which made it easy to lose track of the bigger picture.

A 1 million token context window changes that workflow. A developer could, in principle, give the model a large part of a codebase and ask it to find a bug, add a feature or explain how different parts connect. The maximum output of 128,000 tokens also matters, because long coding jobs need long answers: a full new module, a set of tests or a detailed refactor, not just a few lines.

Large context does not remove the need for human checking. Models can still misunderstand a project, invent a function that does not exist or suggest a fix that looks right but fails in testing. The longer the input, the more important it is to run the code, read the changes and test edge cases before shipping anything.

Visual Understanding: More Than Reading Pictures

LongCat-2.5 is described as multimodal, which means it can work with more than plain text. Visual understanding in this context covers a wide range of jobs: reading charts, summarising a scanned document, describing what is in a photo, checking a design mockup or spotting a problem in a diagram.

That skill is becoming central to everyday AI use. People do not only write prompts in text. They upload screenshots, PDFs, photos and slides and expect the model to make sense of them. A model that handles visual material well can save a lot of manual copying and describing, whether the user is a developer reviewing a user interface or a student working through a textbook diagram.

As with coding, visual answers still need checking. A model can misread a small label, miss a detail in a crowded image or describe something with more confidence than the image supports. For important decisions, the original image or document should always be the final source of truth.

Pricing and Access in Plain Terms

Meituan has priced LongCat-2.5 through its API at 0.30 dollars per million input tokens and 1.20 dollars per million output tokens. Input tokens are what the user sends in, and output tokens are what the model generates. Output costs more because generating text takes more work than reading it.

Those prices put LongCat-2.5 in the competitive range for large models, especially for developers who need long context. A small question will cost only a fraction of a cent, while a huge codebase analysis will cost more, simply because it uses far more tokens. Developers who use the model heavily will need to watch token use, cache repeated material where possible and choose shorter prompts when a short answer will do.

Access starts through Meituan online chat and its API. Releases on GitHub and Hugging Face are planned, which would let developers download and study the model weights and run or adapt the model in their own setups, subject to the licence Meituan publishes. Until those releases appear, the API and chat routes are the confirmed ways to try it.

How LongCat-2.5 Fits the Open Model Race

The past year has seen a rapid rise in large open and open weight models from Chinese developers, alongside open work from US and European labs. These models compete on size, speed, coding skill, language coverage and price, and many are released with the aim of attracting developers who will build products on top of them.

Meituan entry is notable because the company is not primarily known as an AI research lab in the public mind. Its release suggests that large model work is now spreading beyond the first wave of AI focused firms into big platform companies with huge engineering teams and real world data from services, logistics and commerce.

For readers of The Akron Magazine, the key point is choice. More strong open models mean more options for developers, lower prices for some jobs and faster progress on coding assistants and visual tools. It also means more homework for users: comparing models on real tasks, not just on headline numbers, before picking one for serious work.

FAQs

What is LongCat-2.5?

LongCat-2.5 is a large multimodal AI model from Meituan, released on November 2, 2026. It is designed for extended coding tasks and visual understanding, with a 1 million token context window and a maximum output of 128,000 tokens.

How big is the model?

It has 1.6 trillion total parameters, with about 48 billion active per token. That design lets it be very large overall while only using a relevant slice of its capacity for each piece of text it processes.

How much does it cost to use?

Through the API, pricing is 0.30 dollars per million input tokens and 1.20 dollars per million output tokens. Actual cost depends on how much text and code you send in and how long the answers are.

Where can developers get it?

It is available through Meituan online chat and API now, with GitHub and Hugging Face releases planned. Developers should check Meituan official pages for the release status and licence terms before building on it.

Final Thoughts

LongCat-2.5 is a statement of intent as much as a product release. Its size, long context and competitive pricing show Meituan wants a serious place in the developer AI market. Whether it earns that place will depend on real world coding quality, visual accuracy and how smooth the promised open releases turn out to be. For now, it is one more powerful option in a crowded and fast moving field, and developers have good reason to test it on their own projects before judging it. For background on how large models are priced and used, see Reuters on artificial intelligence.