Meituan released LongCat-2.5 model weights under the MIT License, giving developers the option to run or adapt them subject to the license terms. Meituan also offers API access with up to one million tokens of context and temporary usage-based pricing.
Key takeaways
- Meituan published LongCat-2.5 weights on Hugging Face under the MIT License.
- The model is designed for visual input and extended, multi-step tasks such as screenshot inspection and coding work.
- Meituan's API supports up to one million tokens of context and up to 256,000 output tokens.
- Meituan describes the model as having 1.6 trillion parameters, with roughly 48 billion active per token.
- Developers should review the model card, license, infrastructure needs, current pricing, and performance on representative tasks before adopting it.
Meituan has published the open weights for LongCat-2.5, a vision-enabled language model, under the MIT License. Released on Hugging Face on November 9, about a week after the model’s introduction, the weights give developers a way to run and adapt the model under a permissive license, subject to its terms.
LongCat-2.5 is designed for work that combines visual input with extended, multi-step tasks. Meituan also offers API access, with a million-token context window and temporary pricing that makes it possible to evaluate the model without first running its weights.
What Meituan Released and What the MIT License Allows
The release makes LongCat-2.5’s model weights available for developers to use and modify under the MIT License. That distinction matters: open weights allow users to work with the model directly, while access to a hosted API provides a separate way to send requests without managing the model themselves.
The MIT License is permissive, but developers should still review the license and model documentation before deploying the model. They should also check the hardware and infrastructure requirements for running the weights. Publishing weights does not, by itself, mean that a model will be practical to operate on a typical workstation.
LongCat-2.5’s Vision, Coding, and Long-Context Capabilities
Meituan positions LongCat-2.5 for visual tasks and long coding jobs. Its described uses include inspecting screenshots or video, reasoning across a codebase, and making software changes through multiple steps. For example, a developer might ask it to examine an interface screenshot, identify a layout issue, and propose changes to the relevant code. These are intended use cases, not a guarantee that the model will complete every task reliably.
The model’s context window is also a notable part of the offering. Meituan’s API supports up to one million tokens of context and up to 256,000 output tokens. A large context can let a model consider more material in one request - such as extensive code or lengthy task instructions - but the maximum is not necessarily needed for routine work. Developers should test quality, latency, and cost with the amount of context their own tasks require.
Model Scale, API Pricing, and How to Judge Its Positioning
Meituan describes LongCat-2.5 as a 1.6-trillion-parameter model, with roughly 48 billion parameters active per token. These figures indicate a large-scale model with a comparatively smaller subset of parameters active for each token. They do not, on their own, establish how well it performs or how much infrastructure a particular deployment will need.
For API users, Meituan’s limited-time rates are $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens. Actual costs will depend on usage, including how much of a request can use cached input and how long the model’s responses are. Developers should verify the current rates before estimating project costs, since the listed pricing is temporary.
Meituan has used GPT-5.6 Sol and Claude Opus 5 as proprietary reference points in describing the model. Treat those comparisons as selected points of reference, not proof that LongCat-2.5 leads the newest models across every task. Results can vary by benchmark, prompt, and the kind of vision or coding work being tested.
For developers, the release offers two things to evaluate: weights they can run or adapt under the MIT License, and a hosted API with long-context support. Review the model card and license, then test the weights or API on a representative vision or coding task before choosing it for a project.
By the numbers
Meituan describes LongCat-2.5 as a 1.6-trillion-parameter model, with roughly 48 billion parameters active per token.
These figures are reported by Meituan in its description of the model's scale. They do not, by themselves, establish model performance or deployment requirements.
Meituan's API supports up to one million tokens of context and up to 256,000 output tokens.
These are API limits described in the article; the maximum is not necessarily needed for routine tasks.
Meituan lists API rates of $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens.
The article identifies these as limited-time rates from Meituan. Developers should check current pricing before estimating costs.
Meituan published LongCat-2.5's open weights on Hugging Face on November 9, about a week after introducing the model.
The release timing and hosting platform are reported in the article. The weights are provided under the MIT License, subject to its terms.
Step by step
- 01
Review the model card and license
Check the LongCat-2.5 documentation and MIT License to understand the permitted uses, conditions, and available model details before deployment.
- 02
Choose weights or API access
Use the published weights if you need to run or adapt the model directly, or evaluate Meituan's hosted API if you want to test it without managing model infrastructure.
- 03
Check infrastructure requirements
Assess the hardware and operational resources needed to run the weights; do not assume the model will be practical on a typical workstation.
- 04
Test a representative task
Evaluate the model on your own visual or coding workflow, such as interpreting a screenshot or working across a codebase, and measure result quality and latency.
- 05
Estimate usage costs
For API evaluation, model uncached input, cached input, and output token usage separately, then verify Meituan's current rates because the listed prices are temporary.
Frequently asked questions
What is Meituan LongCat-2.5?
Meituan LongCat-2.5 is a vision-enabled language model designed for visual tasks and extended, multi-step work. Meituan describes use cases including inspecting screenshots or video and reasoning across codebases. Its published weights and hosted API provide two ways for developers to evaluate it.
Can developers use LongCat-2.5 under the MIT License?
Yes, Meituan has published LongCat-2.5's weights under the MIT License. Developers should review the license and model documentation before using or deploying the model. The license does not determine whether the hardware and infrastructure requirements are practical for a particular project.
How large is LongCat-2.5's context window?
Meituan says its API supports up to one million tokens of context and up to 256,000 output tokens. The maximum context can help with requests involving extensive code or lengthy instructions, but routine tasks may need much less. Developers should test quality, latency, and cost at the context lengths their workflows require.
How much does the LongCat-2.5 API cost?
Meituan lists temporary rates of $0.30 per million uncached input tokens, $0.006 per million cached input tokens, and $1.20 per million output tokens. Actual costs depend on token use and how much input is served from cache. Verify the current rates before budgeting because the listed pricing is temporary.
