Meituan (LongCat team) · Agentic MoE LLM

LongCat-2.5

LongCat-2.5 is a native multimodal mixture-of-experts model with 1.6 trillion total parameters and about 48 billion activated per token, built for long-horizon agentic tasks such as codebase development and visual understanding. Meituan reports 86% higher per-device throughput than a DDP baseline in long-sequence SFT, and W4A4 quantisation of expert weights with 43.3% higher prefill throughput than a W8A8 baseline. Its evaluation tables compare it with GPT-5.5, Claude Opus 5 and other models, though unmarked scores are in-house and the page calls them indicative. Released under the MIT licence on Hugging Face.