Meituan’s LongCat team launched LongCat-2.5-Preview on its own API platform on September 25. It’s the next generation after LongCat-2.0, which was open-sourced in late June, and the official pitch boils it down to two words: long-horizon tasks and multimodal.
A day later, the overseas open-source coding tool OpenCode announced it would list the model for free, open for two weeks, with zero data retention noted.
Same parameter count, now with eyes
LongCat-2.5-Preview sticks with the mixture-of-experts (MoE) approach, with total parameters around 1.6 trillion and roughly 48 billion activated per inference — the same tier as the 2.0 release. It natively supports a 1-million-token context window, with a maximum single output of 128K tokens.
The changes cluster in two areas. First, image understanding is new: the model can parse image content for cross-modal Q&A, content summarization, and visual reasoning. Second, it’s built for "long-workflow" tasks — Meituan lists scenarios including terminals, browsers, desktop software, spreadsheets, and design tools, meaning the model can operate across multiple interfaces continuously within a single task.
Coding is the centerpiece of this launch. Meituan says the model stands out at code generation, code comprehension, and automated programming tasks, and it’s compatible with mainstream coding tools including Claude Code, OpenCode, OpenClaw, Hermes, and Kilo Code. The API supports both OpenAI and Anthropic formats.
Meituan hasn’t published any benchmark scores so far, so there’s no comparable data yet on how much it has improved over 2.0 on GUI operation and long-workflow tasks. This release is labeled "Preview," and the company hasn’t said whether it will open-source the weights the way it did with 2.0.
What it means for developers in China
The most immediate benefit is low switching cost. Because the API is compatible with the Anthropic format, developers used to Claude Code can swap the backend to LongCat just by changing an API endpoint and key, without rewriting their workflow. For teams in China where calling overseas models isn’t convenient, that amounts to a ready-made alternative backend.
On pricing, Meituan has handed existing users a 5-million-token free quota. Official pricing hasn’t been announced yet, and whether the Preview-phase pricing will match the eventual production pricing also remains to be confirmed by the company.
The combination of long context and long workflows is aimed at codebase-level work. A million tokens can roughly hold the main source code of a mid-sized project, letting the model read it all before making changes — saving a lot of the hassle of stitching context together from fragments. How many steps a multi-step operation can sustain before drifting off track will only become clear once developers actually put it through real use.
From Owl Alpha to 2.5
LongCat’s path so far has a discernible pattern. Before the 2.0 release was open-sourced, it was quietly offered for free on OpenRouter under the codename Owl Alpha, and usage briefly topped the platform’s charts before its identity was revealed and it was open-sourced under the MIT license; Meituan says both training and inference run on domestic chip clusters.
With 2.5-Preview, the company has again chosen to go free first and figure out pricing later: a trial quota on its own platform, and a two-week free period through an overseas tool. Real-world usage volume and developer feedback from those two weeks on OpenCode will likely shape the eventual pricing.
Sources: Meituan LongCat API platform, ITHome, OpenCode’s official account, CocoLoop; verified: ~1.6 trillion total parameters, ~48 billion activated parameters, 1-million-token context window, 128K max output, and the 5-million-token trial quota.