Meituan Open-Sources LongCat-2.5, 1.6T MoE, Under MIT License

Meituan, the Chinese food delivery and e-commerce company, published the full weights of LongCat-2.5 on Hugging Face on November 9, one week after introducing the model. The license is MIT, permitting commercial use and modification. The checkpoint is 1.8TB and carries 1.6 trillion total parameters with roughly 48 billion active per token under mixture-of-experts routing, according to VentureBeat.

The benchmark gains over Meituan's prior release are the clearest selling point. LongCat-2.5 "scores 88.4 versus 70.8 on Terminal Bench v2.1, 64.6 versus 9.7 on DeepSWE v1.1, and 58.3 versus 33.1 on the SWE Atlas codebase question-answering test," VentureBeat reports. Against current frontier closed models the picture is tighter: within one point of GPT-5.6 Sol and Claude Opus 5 on Terminal Bench v2.1, behind Claude Opus 5 on the v4.0 variant at 38.5 to 53.9, and behind DeepSeek-V4.1-Flash on DeepSWE at 74.2. Vision2Web lands at 78.6, 3DCodeBench at 55.4.

Hosted inference during a limited-time launch runs $0.30 per million uncached input tokens and $1.20 per million output. Cached input drops to $0.006. Self-hosters get a different path: Meituan points to an "SGLang-FluentLLM deployment guide for neural processing units," with four-bit precision optimization for expert weights aimed at Chinese accelerators.

The catch is political. VentureBeat flags growing scrutiny in Washington over Chinese AI models and warns that "suppliers should check their contract terms...before putting it into government-facing work." That places LongCat-2.5 inside a dense run of coverage: it is the latest of 304 China AI stories we have tracked in the past ninety days, many of them open-weights releases where the compute is cheap and the procurement question is what the buyer ends up pricing in.