November 2, 2026, (Inside AI) — Meituan released LongCat-2.5 on Monday, a multimodal AI model with 1.6 trillion total parameters that activates roughly 48 billion parameters per token. The company says the model is built for tasks that require AI agents to work continuously over long periods, particularly on large codebases and full software projects.
The launch puts Meituan, best known for food delivery and local services in China, deeper into the frontier model race. LongCat-2.5 adds stronger visual understanding and coding capabilities to the company's existing model lineup. It is available now through Meituan's online chat service and API platform, with releases on GitHub and Hugging Face planned to follow.
The model's design reflects a shift in how developers evaluate AI systems. Parameter count alone no longer tells the story. LongCat-2.5 uses a sparse architecture, activating only a fraction of its total size for each token. This keeps inference costs manageable while preserving a very large knowledge base. The approach mirrors what other major labs have done with mixture-of-experts models, though Meituan's total parameter figure sits at the high end of publicly disclosed systems.
Long Tasks Expose Where Models Break
Meituan said it expanded training on long task sequences and used large-scale reinforcement learning to improve performance on complex codebases. The company also converts failures encountered by real agents into reproducible training tasks. That detail matters. Agentic coding systems often fail not on a single line of code but after dozens of steps, when small errors compound. By turning real-world agent failures into training data, Meituan is targeting the gap between demo performance and production reliability.
Read: Factory Raises $200M, Triples Valuation to $5 Billion
The model processes visual and text inputs within a unified token framework. That means it can handle visual coding, creative design, and 3D tasks without separate pipelines for each input type. For developers building tools that mix screenshots, diagrams, and source code, a single framework reduces integration work.
Meituan's move follows a broader pattern among Chinese technology companies. Alibaba, ByteDance, and Baidu have all released models aimed at coding and agentic workflows in the past year. Meituan's entry differs in its emphasis on sustained execution rather than raw benchmark scores. The company is positioning LongCat-2.5 as a workhorse for agents that run for hours, not minutes.
The competitive logic is straightforward. Meituan operates one of China's largest digital platforms, with millions of merchants and delivery partners. Internal demand for automation is substantial. A model tuned for long coding tasks could support internal engineering teams and eventually be sold to external developers through the API. That dual-use strategy is common among Chinese platform companies that build AI infrastructure for their own operations before commercializing it.
The sparse activation design also carries trade-offs. A model that activates 48 billion parameters per token can be cheaper to serve than a dense model of similar total size. But routing decisions inside the model add complexity. Poor routing can degrade quality on specialized tasks. Meituan has not disclosed benchmark results, so independent evaluation will be needed before developers can judge whether LongCat-2.5 delivers on its coding claims.
The planned GitHub and Hugging Face releases signal that Meituan wants community adoption, not just API revenue. Open-weight releases have become a standard playbook for Chinese AI labs seeking global developer mindshare. The strategy lowers barriers for researchers and startups while creating pressure on closed-model providers to justify their pricing.
Meituan has not announced pricing for API access or a timeline for the open releases. The company also has not published independent benchmark comparisons against competing models. Those details will determine whether LongCat-2.5 becomes a practical tool for developers or remains a showcase project.
Read: DeepSeek Unveils DSec Sandbox Infrastructure for Large-Scale Agent Training
For now, the launch adds another serious contender to the market for agentic coding models. The focus on long-horizon tasks and real failure data suggests Meituan understands where current systems fall short. Whether the model closes that gap will be tested by developers in the coming months, not by press releases.
