November 17, 2026

Meituan’s 1.6T-Parameter LongCat-2.5 Trained Entirely on Chinese Chips

Meituan has open-sourced LongCat-2.5, a 1.6-trillion-parameter agent model trained and deployed exclusively on Huawei Ascend hardware, rivaling top Western coding models.
Meituan’s 1.6T-Parameter LongCat-2.5 Trained Entirely on Chinese Chips

Chinese food delivery company Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter agent model built for long, multi-step tasks such as coding, according to a report from the-decoder.com.

The model is a mixture-of-experts system that activates roughly 48 billion parameters per token, handles up to one million tokens of context, and generates responses up to 128,000 tokens long. It processes text, images and video, and unlike many multimodal systems it converts images into tokens processed alongside text rather than relying on a bolted-on image encoder, an approach drawn from Meituan’s research model LongCat-Next.

For training, Meituan’s LongCat team used reinforcement learning on real GitHub pull request tasks and recreated agent failures, blocking internet access and stripping cached files to prevent the model from gaming its training environment. The company also distilled capabilities from several specialized teacher models into LongCat-2.5.

According to the-decoder.com, Meituan quantizes expert weights to 4-bit precision, which Chinese chips handle natively, boosting throughput by more than 40 percent over an 8-bit version while maintaining accuracy. The self-hosting guide calls for a cluster of 194 Huawei Ascend Atlas A2 accelerators, with no instructions provided for Nvidia GPUs.

On benchmarks, LongCat-2.5 scored 88.4 on Terminal Bench v2.1, close to Claude Opus 5’s 89.1 and GPT-5.6 Sol’s 88.8, with Deepseek-V4.1-Flash topping the table at 90.6. On the harder Terminal Bench v4.0, it scored 38.5, ahead of GPT-5.6 Sol but behind Opus 5. It also posted the highest score in the comparison, 91.9, on the BrowseComp web search benchmark, though Meituan ran that test itself.

The weights are released under the MIT license, and the model is accessible via chat interface and an API compatible with OpenAI and Anthropic formats. Meituan previously trained predecessor LongCat-2.0 on more than 50,000 Chinese AI chips and has also released the efficiency-focused image generator LongCat-Image.

Based on reporting by the-decoder.com.

Leave a Reply

Your email address will not be published. Required fields are marked *