# Meituan releases open weights for 1.6T LongCat-2.5

> The agent model is available under an MIT license and can run on Huawei Ascend chips. Meituan reports strong results on several coding and agent benchmarks.

Oossa · 2026-11-15 · https://oossa.com/en/meituan-releases-open-weights-for-1-6t-longcat-2-5

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model designed to handle long, multi-step tasks, especially coding. The Decoder reports that the model was trained entirely on Chinese hardware and that its weights became available one week after its unveiling.

LongCat-2.5 is a mixture-of-experts model: it has specialized subnetworks, but uses only some of them for each token. Although it contains 1.6 trillion parameters in total, about 48 billion are active per token. Meituan’s API supports up to one million tokens of input and responses up to 128,000 tokens. The model handles text, images and video, and users can switch its reasoning mode on or off.

## What can developers run?

The weights are released under the MIT license, which permits commercial use, modification and redistribution. Developers can also try the model through a chat interface or API. The API supports OpenAI and Anthropic formats and works with agent tools including Claude Code, Codex and Hermes. Meituan’s launch promotion prices input at about €0.27 per million tokens and output at about €1.07 per million tokens.

Meituan has also released a 4-bit version, which reduces the numerical precision used to store model weights and activations. The company says this version keeps accuracy on par with the uncompressed model. In its load tests, it increased prompt-processing throughput by 43.3 percent and response-generation throughput by 42.77 percent compared with an 8-bit version. A self-hosting guide describes a deployment using 194 Huawei Ascend Atlas A2 accelerators: two for image processing, 64 for prompts and 128 for responses.

## How strong are the results?

Meituan reports 88.4 on Terminal Bench v2.1, which tests command-line tasks. That is below Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8, and above Deepseek-V4.1-Flash’s 90.6 only in the sense that it is not: Deepseek leads the listed results at 90.6. On Terminal Bench v4.0, LongCat-2.5 scored 38.5, ahead of GPT-5.6 Sol’s 37.3 but behind Claude Opus 5’s 53.9.

On DeepSWE, a test of fixing bugs in real code repositories, Meituan reports a rise from 9.7 for the earlier result to 64.6 for LongCat-2.5. It remains behind Deepseek-V4.1-Flash at 74.2 and GPT-5.6 Sol at 72.7. The model scored 91.9 on BrowseComp, a web-search benchmark, the highest result in Meituan’s table; the company ran that test in its own environment. These are Meituan’s reported evaluations, and The Decoder does not cite independent benchmark runs.

## The facts

- LongCat-2.5 has 1.6 trillion total parameters, with about 48 billion active for each token.
- Meituan reports 88.4 on Terminal Bench v2.1 and 38.5 on Terminal Bench v4.0.
- The self-hosting guide describes a setup with 194 Huawei Ascend Atlas A2 accelerators.
- The weights use the MIT license, and promotional API prices are about €0.27 per million input tokens and €1.07 per million output tokens.

## Why it matters

For developers who want to adapt or commercially deploy an agent model, the MIT license allows modification and redistribution, and the API offers a way to try LongCat-2.5 without setting up its hardware. Self-hosting, however, is documented on a 194-chip Huawei cluster; the reported benchmark results are Meituan’s own, so how they compare in independent tests remains unknown.

## Sources & references

1. [Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU](https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/) – The Decoder, 2026-11-15

Last updated: 2026-11-15