LongCat-2.5 is Meituan’s new preview model, and it is huge: about 1.6 trillion parameters, a 1 million token context window, and a focus on long, multi-step agent tasks. You can try it now through Meituan’s chat and API. Weights and benchmarks are still missing, so treat the big claims as early.
Meituan, the Chinese delivery and services giant, announced the preview on X. It says the model is built for “long-horizon tasks” across terminals, browsers, spreadsheets and design tools.
Here are the facts that matter most right now:
- Released as LongCat-2.5-Preview on November 2, 2026, per AI Weekly.
- About 48B of the 1.6T parameters are active at a time.
- Live now via Meituan’s chat and API.
- Open weights are not confirmed yet.
Those numbers come from Meituan’s own announcement, so nobody has independently tested them yet.
What LongCat-2.5 actually is
LongCat-2.5 is a multimodal model. It reads text and images in one system, so you can hand it a screenshot or a design and ask for code or edits. The “preview” label matters: Meituan is testing it in public before a final release.
The size sounds scary, but only about 48 billion parameters work on each token. That is the usual trick with large mixture-style models: a huge pool of knowledge, a small slice of compute per step. It keeps answers faster and cheaper than a dense 1.6T model would be.
| Spec | LongCat-2.5-Preview |
|---|---|
| Total size | About 1.6 trillion parameters |
| Active per token | About 48 billion |
| Context window | 1 million tokens |
| Inputs | Text and images |
| Open weights | Not confirmed yet |
The table shows what Meituan has stated so far. Anything else, such as exact architecture details, is not confirmed.
Why the 1M token window matters for agents
A 1 million token window can hold a large codebase, a long contract pile, or hours of agent logs at once. Agents fail most often when they lose track of earlier steps. More memory helps, though it is no guarantee.
That is why Meituan pitches this model at long tasks: running terminal commands, clicking through browsers, filling spreadsheets. If you build coding agents, you will care about this more than any leaderboard number.
Keep one caution in mind. A big window is not the same as using it well. Many models get worse at the far end of a long context, and only independent tests will show whether LongCat-2.5 holds up.
What we don’t know yet
Meituan has not published a full technical report with benchmark numbers we could check. Coding and agent claims come from the company itself. Until outside testers run it, treat any “beats GPT or Claude” headline as marketing.
Open weights are also unclear. Some outlets say GitHub and Hugging Face releases are planned, but no license or date is confirmed. Pricing was not clearly stated in the sources we found, and one outlet’s figures conflicted with the others, so check Meituan’s own API page before budgeting.
Training hardware is another open point. Crypto Briefing reports it was trained on domestically produced Chinese accelerators, but no chip family or cluster size has been disclosed.
How it fits with other big open models
LongCat-2.5 lands in a busy month for large models. If you want a comparison point for sheer size, our look at Mistral Large 4’s open weights covers another trillion-parameter release you can actually download.
For coding agents specifically, see DeepSeek’s free coding agent. And to see why one score never tells the whole story, read why the best AI model still scores 12% on GDPx.
The pattern is clear: Chinese labs are shipping large agent-focused models fast, often with cheap API access. That puts price pressure on the big US labs.
How to test LongCat-2.5 without wasting a day
Start small. Pick one real task you already know the answer to, such as fixing a bug in a repo you wrote or cleaning a messy spreadsheet. Compare the result with what your current tool gives you.
Then stretch it. Feed the model a long file set and ask a question that depends on a detail buried near the middle. That tells you far more about a 1M token window than any marketing chart.
Finally, keep private data out of it. This is a preview served from a company in another country, and you should read the terms and privacy policy before pasting in work files or customer details.
Frequently Asked Questions
What is LongCat-2.5?
It is a multimodal AI model from Meituan with about 1.6 trillion parameters and a 1 million token context window. It is aimed at long agent tasks like coding, browsing and spreadsheets.
Can I use LongCat-2.5 today?
Yes, as a preview. Meituan says it is live through its chat service and API. Availability and sign-up rules may vary by region.
Is LongCat-2.5 open source?
Not yet. Some reports say weights are planned, but no release date or license has been confirmed.
Is LongCat-2.5 better than Claude or GPT?
Nobody knows yet. Meituan has not published checkable benchmarks, so wait for independent tests.
Should you care right now?
If you build AI agents, spend an hour testing the preview on your own tasks, using throwaway data only. If you are a regular user, there is nothing to switch to yet. Wait for open weights, a license and outside benchmarks before you rely on LongCat-2.5 for real work.
Leave a Reply