Monday, 16 November 2026, story 1 of 4
Meituan publishes weights for its 1.6-trillion-parameter LongCat-2.5
The MIT-licensed agent model trails Claude Opus 5 on Meituan's harder terminal tests, and its self-hosting guide covers only Huawei Ascend chips.

What happened
Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model built for long, multi-step agent work, particularly coding. The release comes one week after the official unveiling. About 48 billion parameters are active per token. The weights carry an MIT license, which allows commercial use. A chat interface and an API in OpenAI and Anthropic formats are live. Meituan's launch price is a limited-time promotion of roughly €0.27 per million input tokens and €1.07 per million output tokens. The Decoder reports the model was trained entirely on Chinese hardware.
Meituan reports 88.4 on Terminal Bench v2.1. In its table Claude Opus 5 scores 89.1, GPT-5.6 Sol 88.8 and Deepseek-V4.1-Flash 90.6.
What it means
On the easier terminal benchmark, the gap to the leaders is small. On Meituan's own harder numbers it widens: 38.5 on Terminal Bench v4.0 against 53.9 for Opus 5, and 64.6 on DeepSWE, behind every model in the table except Claude Opus 4.8. The table also omits Claude Opus 5.5 and GPT-6.1 Sol. All figures are Meituan's. "Near-frontier" holds for some tasks, not all.
Self-hosting is less open than the license suggests. The only documented deployment is a cluster of 194 Huawei Ascend accelerators, with no guidance for Nvidia GPUs. Most teams will likely meet the model through the API first. Its compatibility with existing agent harnesses makes a trial cheap, but the price is promotional and may change. Independent evaluations are the next thing to watch.
What to do: Point an existing agent harness at the LongCat-2.5 API this week and test it on your own tasks before trusting Meituan's benchmark table.
Source
Filed under Open source. Published .
"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.