Meituan releases LongCat-2.5, a 1.6-trillion-parameter native multimodal model

Meituan's LongCat team released LongCat-2.5 on November 2, a new flagship model with 1.6 trillion total parameters that the company presents as natively multimodal. Rather than leading with benchmark scores, the release and the coverage around it focus on what the model can carry through end to end.

Three capability areas are highlighted. LongCat-2.5 is pitched at long-horizon agent tasks that require planning and sustained execution; it adds screenshot-to-code ability, turning interface captures or design drafts into working code; and it is positioned to support complex, multi-step project development.

The phrase domestic chips, domestic model has become the release's signature in Chinese coverage. Reports describe LongCat-2.5 as leaning on domestic compute and a domestic model stack for training and inference, presented as progress in hardware-software co-design. For teams constrained by access to high-end accelerators, that framing matters more than any single leaderboard result.

The business angle is equally pointed. LongCat-2.5 explicitly strengthens business analysis, a capability that maps onto Meituan's core footprint in local services, supply chains and merchant tooling. That suggests the model is intended first as infrastructure for internal efficiency and merchant-facing products rather than as a consumer chatbot.

LongCat has been visible to developers since its preview releases, and the 2.5 launch moves the line from capability demos to something Meituan can actually ship. The short gap between preview and general release also tracks the accelerating cadence among China's top model teams.

The competitive field has shifted from parameter counts to deployment. Alibaba, ByteDance, DeepSeek and Moonshot have each staked out open-source ecosystems, multimodality, inference efficiency and long context; Meituan's combination of native multimodality, long-horizon agents and vertical business scenarios is narrower but easier to tie to measurable returns.

Three things are worth watching next: whether LongCat-2.5 opens an API or releases weights, whether third parties can reproduce its results on agent benchmarks, and whether Meituan embeds it in merchant back offices and local-services workflows in a way that produces public efficiency numbers.

The real signal here is not the parameter count. It is that a major Chinese platform company is now treating native multimodal agents under compute constraints as an engineering problem to be solved in production, not a demo to be shown.