Meituan builds 1 big AI on Chinese chips

Chinese food delivery giant Meituan released a new AI model, LongCat-2.5, on the 2nd of November 2026. Meituan said it was trained fully on Chinese made computer chips.

The model has 1.6 trillion parameters, the settings an AI learns. It uses about 48 billion of them for each word piece, the LongCat blog said.

Built for long coding jobs

LongCat-2.5 is made for long tasks such as building whole software projects, Meituan said. It can also read pictures and write code from them.

During training it read runs of text up to 1 million tokens long, the company said. Tokens are the word pieces AI reads.

Chinese chips

Each Chinese chip has less than 80 GB of memory, the blog said. Meituan changed how work is split across chips to fit that limit.

Meituan did not name the chip maker. Its earlier model, LongCat-2.0, was also trained on Chinese chips in June.

How it scores

On a coding test called Terminal Bench 4.0, LongCat-2.5 scored 38.5, Meituan said. That beat OpenAI's GPT-5.6 Sol at 37.3 but trailed Anthropic's Claude Opus 5 at 53.9.

These scores come from Meituan's own blog. No outside group has checked them yet.

Below we break down the numbers

About 48 billion out of 1.6 trillion parameters work at once. That is 3 percent of the model, by our arithmetic.

Claude Opus 5 leads LongCat-2.5 by 15.4 points on Terminal Bench 4.0. LongCat-2.0 scored 3 on the same test.

What is confirmed and what is not

The size, the training on Chinese chips and the scores come from Meituan. The chip maker and any price were not given.

Key facts

Meituan released LongCat-2.5 on the 2nd of November 2026.

The model has 1.6 trillion parameters.

Meituan said it was trained on Chinese made chips.

The test scores are Meituan's own.