Meituan launches 1.6-trillion-parameter LongCat-2.5 AI model

Meituan unveiled LongCat-2.5, a multimodal AI model with 1.6 trillion total parameters, and began API service for it on November 2, according to aitimes.com. The weights of the model have not been released.

The model uses a mixture-of-experts architecture that activates about 48 billion parameters per token. Meituan did not disclose the manufacturer or product name of the Chinese-made AI accelerators used to train it.

LongCat-2.5 applies W4A4 quantization, compressing weights and compute data to 4 bits each. Meituan says input throughput rose 43.3% and output generation throughput rose 42.77% compared with the 8-bit method.

Benchmark results

On BrowseComp, a web-search problem-solving benchmark, LongCat-2.5 scored 91.9, up from 79.9 for the previous version and ahead of Claude Opus 5 at 90.8 and GPT-5.6 Sol at 90.4. It scored 55.4 on 3DCodeBench, the highest among the compared models. On DeepSWE 1.1 it reached 64.6, up from 9.7.

On the coding benchmark TerminalBench 4.0, LongCat-2.5 scored 38.5, up from 3.0 for LongCat-2.0, but below Claude Opus 5's 53.9.

LongCat-2.5 is available via web and API, with 10 million free tokens for new sign-ups and 5 million extra for existing users. Meituan trained the previous version, LongCat-2.0, on Chinese-made AI semiconductors in June; the two models share the same model scale.