Chinese food delivery giant Meituan has publicly released the weights for LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model designed for multi-step AI agent tasks and visual coding. Trained entirely on a cluster of 194 Huawei Ascend Atlas A2 accelerators without Nvidia hardware, the model activates 48 billion parameters per token and supports a 1-million-token context window with up to 128,000 generation tokens.

To optimize for non-Nvidia hardware, Meituan native-quantized the model’s expert weights and activations to 4-bit precision, boosting prompt processing throughput by 43.3% and response generation by 42.77% in load tests. The training process leveraged GitHub pull requests, distilled specialized teacher models, and utilized internet-isolated environments to evaluate agent task execution accurately.

On public benchmarks, Meituan reports that LongCat-2.5 achieved 88.4 on Terminal Bench v2.1 and 64.6 on DeepSWE, placing it near leading Western models like Claude Opus 5 and GPT-5.6 Sol. The model’s weights have been published under an open-source MIT license allowing unrestricted commercial use and self-hosting.

Why it matters

  • Proves frontier-class MoE agent models can be trained and run efficiently on non-Nvidia domestic hardware using native 4-bit quantization.

  • Provides developers with an open-source (MIT), highly capable multi-modal agent model featuring a 1M token context window.

  • Demonstrates that RL training using real-world GitHub PRs and execution checking yields strong coding benchmark performance.

Source: the-decoder.com