LongCat-2.5: A Free Chinese AI Model. Should You Use It?
Meituan released LongCat-2.5 as open weights under an MIT license on November 9, 2026. What it is, what it costs, and the data question to ask first.
By Peter Walters
Meituan, the Chinese delivery and services company, released the open weights of its LongCat-2.5 model on Hugging Face on November 9, 2026, under the permissive MIT license. LongCat-2.5 is a mixture-of-experts model with about 1.6 trillion total parameters and about 48 billion active per token, it handles images and video as well as code, and Meituan's hosted API is priced at $0.30 per million input tokens and $1.20 per million output tokens for a limited time.
What actually changed
Open weights means anyone can download the model files, run them on their own hardware and build on them. The MIT license allows commercial use, modification and redistribution, with notice requirements. VentureBeat reports the weights arrived one week after Meituan introduced the model.
The headline scores come from Meituan itself. It reports 88.4 on Terminal Bench v2.1 (its predecessor scored 70.8) and 64.6 on DeepSWE v1.1 (up from 9.7). On Terminal Bench v4.0 it scores 38.5 against 53.9 for Claude Opus 5, and on Vision2Web it scores 78.6 against 80.2 for Opus 5. VentureBeat notes Meituan chose older proprietary models as baselines and that test setups differ, so these numbers do not give a clean head-to-head ranking.
The price is the real story. VentureBeat estimates a job with 200,000 input tokens and 100,000 output tokens costs about $0.18 on LongCat-2.5's hosted API. Its estimate for the same job is $2.80 on Claude Opus 5.5 and $5.60 on Claude Fable 5.2. AI Weekly lists a 1 million token context window.
Running it yourself is not a weekend project. The checkpoint is roughly 1.8 terabytes, and Meituan's deployment guide targets specialized accelerators. This is a model for companies with an infrastructure team, not for a studio with a laptop.
Why it matters if you run a small team
Most of you will never download 1.8 terabytes. What matters is that a capable model now costs a few cents per big job, and that pressure tends to push prices down everywhere. If you pay for AI through a tool, expect that tool to get cheaper or better over the next year.
What is overhyped: the benchmark table. A vendor choosing its own baselines proves very little. A model that wins on coding tests can still be worse at the things you need, like writing a client email in your voice or reading a messy spreadsheet.
What deserves a pause is the data question. VentureBeat advises confirming the hosted service's data retention, access and deletion terms before sending proprietary code or screenshots. I agree, and I would go further: do not send customer lists, financial records or anything a client told you in confidence to any hosted model until you have read where that data goes. That rule applies to every vendor, not just this one. VentureBeat also found no public U.S. federal restriction naming Meituan as of November 9, and says rules can change quickly.
What I would do with it this week
Nothing dramatic. Pick one boring, low-risk task you already do with AI, like summarizing public web pages or drafting a first-pass product description. Run the same prompt through the model you use now and through a cheaper one, then compare the results side by side. If the cheap one is as good, you just cut a cost. If it is not, you learned that in ten minutes.
Then write down which data you will never put into a hosted model, and share the list with your team. When I built a daily brief agent for a real estate agent I work with, the first decision was not which model to use. It was which information the agent was allowed to see. The model choice was easy to change later, and the data boundary was not.
If you are comparing options and want a second set of eyes, start here.
Sources
- VentureBeat: Meituan releases the open weights of its LongCat-2.5 LLM with vision under a permissive MIT license
- AI Weekly: AI News Today, November 10
Want one practical AI move like this in your inbox each week? Join the letter.