LongCat-2.5 activates about 48 billion of its 1.6 trillion parameters per token. Through the API it accepts up to one million tokens at once and generates responses up to 128,000 tokens long. It takes text, images and video as input, replies in text, and lets users switch reasoning on and off.
For reinforcement learning, Meituan used real GitHub pull request tasks and recreated agent failures. To keep the model from taking shortcuts, training sandboxes had no internet access and no Git history, and a language model checked solutions for tricks. The self-hosting guide describes a 194-chip Huawei Ascend cluster and gives no guidance for Nvidia GPUs.
On Terminal Bench v2.1 the model scores 88.4, just below Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8. On DeepSWE it reaches 64.6, behind several models in Meituan's table. It scores 91.9 on BrowseComp, but Meituan ran that test in its own environment. The weights are under the MIT license, and the API follows both the OpenAI and Anthropic formats.
Why it matters if you build agents
Developers can self-host the MIT-licensed weights, but the documented deployment targets Huawei Ascend chips rather than Nvidia GPUs. The API's OpenAI and Anthropic compatibility means it can be plugged into harnesses such as Claude Code and Codex.
The benchmark numbers come from Meituan's own evaluations, and the promotional pricing is described as limited-time, so teams should run their own tests before switching.
Get stories like this every weekday
One short email with the day's agent tooling news.