Meituan released LongCat-2.5 on November 2, adding a 1.6-trillion-parameter native multimodal model to its LongCat lineup — and designating long-horizon agentic coding, not chat, as its primary mission. According to TechNode, reporting on the launch via IT Home, the model activates approximately 48 billion parameters for each token, an architecture designed to deliver frontier-scale capacity without paying the full inference cost of the total parameter count on every request.
The launch is the production follow-up to LongCat-2.5-Preview, which Meituan shipped in late September as a 1.6T mixture-of-experts agent model with a 1 million-token context window, per Pandaily's coverage at the time. For more on this story and the wider model race, see our AI models coverage.
Built for Agents That Work for Hours, Not Seconds
The defining bet of LongCat-2.5 is that the frontier has moved from answering questions to completing tasks. Meituan said it expanded training on long task sequences and applied large-scale reinforcement learning to improve the model's work on complex codebases and complete software projects — the extended, multi-step jobs where agent systems tend to drift, stall or quietly fail.
One training technique stands out. The company said it converts failures encountered by real agents into reproducible training tasks, closing the loop between what production agents get wrong and what the model learns next. It is an approach that treats agent failures not as anecdotes but as curriculum — and it signals that Meituan is harvesting hard training data from deployed systems at a scale most labs cannot match, given the company's food-delivery and logistics operations generate enormous volumes of real-world task traffic.
Native Multimodality in a Unified Token Framework
LongCat-2.5 processes visual and text inputs within a unified token framework, according to the launch coverage — meaning images and text are converted into the same underlying representation rather than patched together through separate encoders. Meituan says this enables visual coding, creative design and 3D tasks alongside conventional software engineering work.
The multimodal emphasis is a differentiator in the agentic-coding niche, where most competitors remain text-first. An agent that can read a design mockup, reason about a 3D scene, or interpret a screenshot of a bug report can take on a wider share of real product-development work than one that only sees code.
The Economics of 48 Billion Activated Parameters
The 1.6T-total / ~48B-active split is worth unpacking, because it explains who this model is for. In a mixture-of-experts design, only a small fraction of the network fires on any given token, which is what makes a trillion-scale model affordable to run: capacity is there for the requests that need it, while the per-token cost tracks the active parameter count. For agentic workloads the math compounds. An agent that works for hours generates hundreds of thousands or millions of tokens across planning, tool calls and code edits — so the gap between paying for 48 billion active parameters and paying for 1.6 trillion on every step is the difference between a viable product and a research demo.
Combined with the 1 million-token context window the preview model introduced, per Pandaily, the target workload is unambiguous: an agent holding an entire codebase, its documentation and its failure history in working memory while it executes a task list measured in hours.
Availability: Chat and API Now, Open Weights to Follow
The model is available immediately through Meituan's online chat service and API platform, according to TechNode. GitHub and Hugging Face releases are planned to follow — a sequencing that echoes the playbook of other Chinese labs, which increasingly ship hosted access first and open weights shortly after.
The staged release matters for two audiences. For enterprises, API availability today means LongCat-2.5 can be evaluated immediately for coding-agent workloads. For the open-source ecosystem, the promised weights release will determine whether LongCat-2.5 becomes a foundation for fine-tuned agentic systems the way previous openly released Chinese models have.
A Crowded Week for Chinese Model Releases
The launch lands amid a burst of activity from China's AI sector. On the same day, TechNode reported that SenseTime open-sourced a 9-billion-parameter model that reuses layers for deeper processing, and that DeepSeek's team open-sourced a proactive coding agent research prototype. Meituan, best known globally as a food-delivery giant, has quietly built one of the more credible AI research programs in the country — and LongCat-2.5 is its argument that agentic coding, multimodal from the ground up, is the workload that matters.
Whether the full release outperforms the preview model on the benchmarks that agencies and developers care about — long-horizon task completion, codebase-level edits, multimodal grounding — will be the immediate test. What is already clear is the strategic direction: the next contested ground for large models is not trivia answers or essay writing, but hours-long autonomous work on real software.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
