最新模型
LongCat-2.5-Preview Adds Image Understanding; OpenCode Offers Limited-Time Free Access
Meituan’s new preview model retains a one-million-token context window and adds image understanding and coding-agent integrations. Official documentation lists inference controls, but the visual input format and performance on real tasks remain to be verified.
On September 25, Meituan launched LongCat-2.5-Preview on its API platform, adding image understanding for multimodal question answering, summarization, and visual reasoning. It also lists integrations with coding agents such as Claude Code, Hermes, and OpenCode. This gives developers a chance to incorporate screen information into code-editing workflows. However, the release notes include no capability comparison scores, so they do not establish success rates for visual debugging or long-running tasks. [Official changelog](https://longcat.chat/platform/docs/ChangeLog.html)
The specifications retain the previous model’s one-million-token context window and support for up to 128K output tokens, so this release should not be interpreted as another expansion of context capacity. The service provides interfaces compatible with OpenAI Chat Completions and Anthropic Messages, allowing it to be connected to existing agent frameworks. In engineering terms, this makes it possible to evaluate the new model within existing tool workflows, but a large context window alone does not guarantee that the model will accurately use every file and item in the conversation history. [Quickstart documentation](https://longcat.chat/platform/docs/)
For inference control, the official documentation says thinking mode is enabled by default and can be set to `enabled` or `disabled` through `thinking.type`. Developers can compare the two modes for code-editing accuracy, response time, and output usage. OpenCode’s integration settings also allow this option to be overridden, making it easier to compare the modes on the same tasks. [OpenCode integration documentation](https://longcat.chat/platform/docs/OpenCode.html)
The model is also listed in OpenCode Zen under the identifier `longcat-2.5-preview-free`, with input, output, and cache reads marked as free. This provides another testing route, but the documentation promises only limited-time availability and gives no end date. Evaluations should record the provider, model identifier, and settings used to avoid mistaking differences between services for changes in model capability when switching endpoints later. [OpenCode Zen documentation](https://opencode.ai/docs/zen/)
The documentation for visual capabilities still has gaps: the changelog announces image support, while the Chat Completions reference page continues to describe message content as a plain-text string and does not show an image request example. The documentation may not yet be synchronized; this alone does not prove that the service lacks image support. Engineering teams should first verify the image format accepted by the endpoint they plan to use, then test screenshots, cross-file edits, and multi-turn tool operations. They should measure task completion rates, retry counts, and total usage. The next items to watch are a complete description of the visual interface and reproducible evaluations. [API reference](https://longcat.chat/platform/docs/api/chat)