Meituan, the Chinese delivery and local services giant, has released the weights for LongCat-2.5, a 1.6 trillion-parameter Mixture-of-Experts model, under an MIT license. The model was unveiled Nov. 1.
The move represents a calculated bet on a different kind of AI strategy. Rather than hoard weights behind an API paywall, Meituan is releasing full model parameters to developers and researchers. The MIT license permits commercial use, modification, and redistribution.
MoE architectures activate only a subset of model parameters per inference—a design that matters economically. A dense 1.6T-parameter model would require loading all weights into memory and computing all of them; MoE pays for compute only on the subset used. For a delivery platform running thousands of simultaneous inferences, that efficiency translates directly to GPU utilization rates and inference margin.
LongCat-2.5 is built for long-horizon agentic tasks—multi-step reasoning over extended contexts. Inside Meituan's ecosystem, that means AI that can handle complex logistics optimization, dynamic routing, or customer support conversations that require memory and sequential problem-solving. The release suggests the company sees this capability as table stakes, not differentiation.
Open-sourcing foundational models serves multiple payoffs for Meituan. It attracts researchers and engineers who want to work on frontier AI; it builds goodwill with the developer community; it positions the company as a steward of AI infrastructure rather than just a consumer. More subtly, it shapes the baseline capabilities that competitors will use. If LongCat becomes a community standard, Meituan's internal finetuned versions (kept private) run on better base models than any closed alternative could offer.
This mirrors the playbook of other infrastructure players. Meta open-sourced Llama; Stability AI released Stable Diffusion weights. Each company retains proprietary applications and finetuning; the open release commoditizes the base layer and shifts competition upstream to domain expertise and integration.
For Meituan specifically, the calculus is straightforward: foundational model development is capital-intensive but fast-commoditizing. The company has the GPU capacity to train 1.6T-parameter models; most competitors don't. Open-sourcing the result costs almost nothing in marginal terms and buys influence. Proprietary lock-in on base models no longer works; the real moat is in the applications and finetuning that sit on top.
