Meituan has released LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts (MoE) agent model trained entirely on domestic Chinese hardware, activating roughly 48 billion parameters per token. Designed for autonomous, multi-step coding tasks, the multimodal model processes text, images, and video natively by converting images into tokens alongside text, supporting context windows up to one million tokens and generating responses up to 128,000 tokens. The model features a toggle for reasoning and was trained using reinforcement learning on real GitHub pull requests and recreated agent failures within isolated environments lacking internet access, Git history, or cached files, with solutions verified by an LLM to prevent reward hacking. Capabilities were further distilled from specialized teacher models. For inference, Meituan quantizes weights and activations to 4-bit precision for native execution on Huawei Ascend Atlas A2 accelerators, reporting a 43.3% throughput increase for prompt processing and 42.77% for generation over an 8-bit baseline without accuracy loss. A reference deployment utilizes 194 chips: two for image processing, 64 for prompt ingestion, and 128 for generation. On Terminal Bench v2.1, LongCat-2.5 scores 88.4, trailing Claude Opus 5 (89.1) and DeepSeek-V4.1-Flash (90.6). On DeepSWE, it reaches 64.6, up from 9.7, but lags behind leaders DeepSeek-V4.1-Flash (74.2) and GPT-5.6 Sol (72.7). It scores 38.5 on the harder Terminal Bench v4.0, above GPT-5.6 Sol (37.3) but below Opus 5 (53.9), while achieving a leading 91.9 on BrowseComp in Meituan's environment. Weights are available under the MIT license, with API access priced promotionally at approximately €0.27 per million input tokens and €1.07 per million output tokens.

Read original