The Pulse
Meituan Opens LongCat-2.5-Preview for Agentic Coding
Meituan’s LongCat API change log lists LongCat-2.5-Preview as available from September 25, with image understanding and compatibility with several coding and agent tools. Its documentation specifies a one-million-token context window, a 128
AI.info Team · Written with AI · fact-checked ·
Meituan made LongCat-2.5-Preview available on September 25, positioning the model for coding and agent tasks. The release adds image understanding to the LongCat API platform and lists compatibility with developer tools including Claude Code, Hermes, OpenClaw, OpenCode and Kilo Code. The change log describes the model as intended for code generation, code understanding and automated programming tasks.
Image input joins the coding pitch
LongCat’s announcement lists image understanding as a new capability, including cross-modal question-answering, content summarization and complex visual reasoning. Developers can use the model’s multimodal capability to work with image content alongside text. The change log describes these capabilities but does not provide benchmark results or comparative performance scores for the preview.
The release is dated September 25 in LongCat’s platform change log. That entry names compatibility with a group of coding and agent tools rather than limiting use to a LongCat application. The documentation gives specific setup instructions for OpenCode and lists Claude Code, Hermes, OpenClaw and Kilo Code among the environments the model works with.
A million-token context window, 128K output cap
LongCat’s quick-start guide specifies a one-million-token context window for LongCat-2.5-Preview and a maximum output length of 128,000 tokens per request. These are separate limits: the context window sets the amount of material available to the model, while the output cap limits what it can generate in one response.
The API documentation describes an OpenAI-compatible chat-completions endpoint that supports both streaming responses, delivered through server-sent events, and non-streaming responses. Requests use a bearer API key and include a model identifier and messages; the documented message roles are system, user and assistant. The endpoint also allows parameters such as temperature and a thinking setting, which can be enabled or disabled in the request.
The platform also supports an Anthropic-compatible API format, with a separate messages endpoint. The quick-start guide identifies both formats as available for LongCat-2.5-Preview. This gives developers a choice of API conventions for connecting to the model, while the OpenCode guide documents a built-in LongCat provider with the endpoint and model list preconfigured.
OpenCode setup and preview status
For OpenCode, LongCat’s instructions direct users to authenticate with the command-line interface, choose LongCat as the provider and enter an API key. After authentication, users can select LongCat-2.5-Preview from the model list. The guide says thinking is enabled by default and can be turned off through the model’s configuration options.
The preview status is relevant to the claims around performance: LongCat describes coding and image-understanding capabilities, but the change-log entry does not include independent tests or comparative evaluations. The documentation’s concrete details are the supported API formats, model limits and integration steps. Developers can test the model in OpenCode or connect through either documented API format, with requests capped at 128,000 output tokens even when the input fits within the one-million-token context window.