Meituan open-sources 1.6T-parameter LongCat-2.5, trained without Nvidia chips

Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model built for long-horizon autonomous coding and agent tasks, trained entirely on Chinese-made accelerators, with a self-hosting guide built on 194 Huawei Ascend chips.
Meituan has released the open weights for LongCat-2.5, a mixture-of-experts (MoE) model with 1.6 trillion parameters, designed for long, multi-step autonomous tasks, especially coding. The headline detail is that it was trained entirely on Chinese-made computing hardware, without Nvidia chips.
The Decoder reports the weights became available about a week after the model was unveiled, under the MIT licence, which permits commercial use, modification and redistribution. Although the model has 1.6 trillion parameters in total, only about 48 billion are active per token. Its API supports up to one million tokens of input and responses up to 128,000 tokens, handles text, images and video, and lets users switch reasoning mode on or off.
Self-hosting guide runs on 194 Huawei Ascend chips
Meituan's self-hosting guide deploys the model on 194 Huawei Ascend Atlas A2 accelerators: two for image processing, 64 for prompts and 128 for responses. The guide makes no mention of Nvidia hardware or deployment. Meituan also released a 4-bit quantised version, which it says keeps accuracy on par with the uncompressed model while increasing prompt-processing throughput by 43.3 percent and response-generation throughput by 42.77 percent over an 8-bit version.
Benchmarks are Meituan's own; no independent verification yet
Meituan reports 88.4 on Terminal-Bench v2.1, close to Claude Opus 5 at 89.1 and below DeepSeek-V4.1-Flash at 90.6. On DeepSWE, a test of fixing bugs in real code repositories, it scored 64.6, a sharp rise from 9.7 for the previous version, but still behind DeepSeek at 74.2. These are the company's own reported evaluations, and no independent benchmark runs have been published yet.
The model's API is compatible with OpenAI and Anthropic formats and works with agent tools including Claude Code and Codex. Launch pricing is about €0.27 per million input tokens and €1.07 per million output tokens.
(Photo: file photo of an AI chip.)
N環球 
💬 Comments
No comments yet. Be the first!