All posts
2 min readby Romiel Inolino

Meituan released LongCat-2.5 under MIT: 1.6T parameters, 88.4 on Terminal Bench v2.1, and an API that speaks both OpenAI and Anthropic formats

AI modelsopen weightsAI agentsLLM APImodel selection

Meituan published the weights for LongCat-2.5, a large open model built for long agent tasks, with a promotional API price far below most frontier models.

What happened

LongCat-2.5 is a mixture-of-experts model with 1.6 trillion total parameters and about 48 billion active per token, according to its Hugging Face model card. The Decoder reports that the API takes up to one million tokens of input, returns up to 128,000 tokens, reads text, images and video, and lets users switch reasoning on and off. The weights are under the MIT license, which allows commercial use.

In Meituan's own benchmark table, LongCat-2.5 scores 88.4 on Terminal Bench v2.1, against 89.1 for Claude Opus 5 and 88.8 for GPT-5.6 Sol. On the harder Terminal Bench v4.0 it scores 38.5, well below Opus 5 at 53.9. The model card says the terminal tests were run through Claude Code. Newer models such as Claude Opus 5.5 and GPT-6.1 Sol aren't in the comparison, as The Decoder notes.

The Decoder reports the API accepts both OpenAI and Anthropic request formats and works with agent tools such as Claude Code and Codex, at a limited-time promotional price of about 0.27 euros per million input tokens and 1.07 euros per million output tokens. Meituan's self-hosting guide describes a cluster of 194 Huawei Ascend accelerators and offers no Nvidia guidance.

My take

For most teams I work with, self-hosting a 1.6 trillion parameter model is not realistic, so the API is the part that matters. Compatibility with both request formats means trying it can be as small as changing a base URL and model name in an existing workflow, which makes a side-by-side test cheap.

I'd treat these scores as a reason to run your own evals, not as a verdict. They come from Meituan, the comparison skips the current top models, and the sources don't say when the promotional price ends. Before routing client data through any new provider, I'd also check where requests are processed and what the data retention terms say, since the sources I read don't cover that.

A reasonable test: take 50 real tasks from an existing agent step, run them on your current model and on LongCat-2.5, and compare success rate, cost per task and latency.

More posts