Meituan Releases LongCat-2.5, Opens API

Meituan said on the 2nd local time that it had released LongCat-2.5, a multimodal model trained on Chinese-made AI chips, and launched API access. The model has 1.6 trillion total parameters and uses a mixture-of-experts (MoE) architecture, activating about 48 billion parameters for each token. Its total size matches the earlier LongCat-2.0, but Meituan says it has improved visual understanding and agentic coding, and supports complex software development, visual design and 3D content creation. The company said the model is much better than its predecessor at carrying out long-running tasks.

In Meituan’s own evaluations, Terminal-Bench 4.0 rose from 3.0 to 38.5, and DeepSWE 1.1 increased from 9.7 to 64.6. BrowseComp, which evaluates problem-solving using web searches, climbed from 79.9 to 91.9. On BrowseComp, LongCat-2.5 scored above Claude Opus 5 at 90.8 and GPT-5.6 Sol at 90.4; it also recorded the highest score among the compared models on 3D-CodeBench, at 55.4. However, it fell short of Claude Opus 5’s 53.9 on Terminal-Bench 4.0. These scores were disclosed by Meituan, and some competitors’ results were measured by the company itself.

Meituan used long-horizon reinforcement learning (training through repeated actions and feedback) to improve performance. It gathered agent failures from real software development and turned them into reproducible training tasks. Using GitHub code-change requests and related issues, it automatically generated training environments and validation tests. The company also applied techniques to reduce repeated reasoning and unnecessary tool calls. It improved multimodal processing by converting visual information into multiple layers of digital tokens that the model handles with the same architecture used for text.

Meituan demonstrated examples of webpage design, 3D models and interactive content, including a webpage with real-time fluid simulation and a digital exhibition page for rotating and zooming in on 3D artifacts. For business tasks, the model supports analysis of business data, calculation of returns, classification of customer feedback and report writing. Meituan said it had trained LongCat-2.0 on Chinese-made AI chips in June and also used domestic computing infrastructure for this model, but did not disclose the chip maker or product.

Meituan optimized its long-context training and inference system for the memory and computing characteristics of Chinese-made accelerators. It applied W4A4 quantization, compressing model weights and computation data to 4 bits each. The company said this increased input throughput by 43.3% and output-generation throughput by 42.77% compared with the existing 8-bit method. LongCat-2.5 is available on the web and through an API, with an extra 10 million tokens for new users and 5 million tokens for existing users. The model weights have not yet been released.

Meituan said on the 2nd local time that it had released LongCat-2.5, a multimodal model trained on Chinese-made AI chips, and launched API access. The model has 1.6 trillion total parameters and uses a mixture-of-experts (MoE) architecture, activating about 48 billion parameters for each token. Its total size matches the earlier LongCat-2.0, but Meituan says it has improved visual understanding and agentic coding, and supports complex software development, visual design and 3D content creation. The company said the model is much better than its predecessor at carrying out long-running tasks.

In Meituan’s own evaluations, Terminal-Bench 4.0 rose from 3.0 to 38.5, and DeepSWE 1.1 increased from 9.7 to 64.6. BrowseComp, which evaluates problem-solving using web searches, climbed from 79.9 to 91.9. On BrowseComp, LongCat-2.5 scored above Claude Opus 5 at 90.8 and GPT-5.6 Sol at 90.4; it also recorded the highest score among the compared models on 3D-CodeBench, at 55.4. However, it fell short of Claude Opus 5’s 53.9 on Terminal-Bench 4.0. These scores were disclosed by Meituan, and some competitors’ results were measured by the company itself.

Meituan used long-horizon reinforcement learning (training through repeated actions and feedback) to improve performance. It gathered agent failures from real software development and turned them into reproducible training tasks. Using GitHub code-change requests and related issues, it automatically generated training environments and validation tests. The company also applied techniques to reduce repeated reasoning and unnecessary tool calls. It improved multimodal processing by converting visual information into multiple layers of digital tokens that the model handles with the same architecture used for text.

Meituan demonstrated examples of webpage design, 3D models and interactive content, including a webpage with real-time fluid simulation and a digital exhibition page for rotating and zooming in on 3D artifacts. For business tasks, the model supports analysis of business data, calculation of returns, classification of customer feedback and report writing. Meituan said it had trained LongCat-2.0 on Chinese-made AI chips in June and also used domestic computing infrastructure for this model, but did not disclose the chip maker or product.

Meituan optimized its long-context training and inference system for the memory and computing characteristics of Chinese-made accelerators. It applied W4A4 quantization, compressing model weights and computation data to 4 bits each. The company said this increased input throughput by 43.3% and output-generation throughput by 42.77% compared with the existing 8-bit method. LongCat-2.5 is available on the web and through an API, with an extra 10 million tokens for new users and 5 million tokens for existing users. The model weights have not yet been released.