MoonshotAI: Kimi K2.5

Kimi K2.5 is a text generation model from Moonshot AI, listed on api14 under the model id moonshotai/kimi-k2.5. It sits in the text category and is described by its provider as a native multimodal model built on Kimi K2, with continued pretraining over roughly 15T mixed data. If you have used earlier Kimi releases, this is the successor to plan around when you need current Moonshot AI behaviour rather than an older checkpoint.

The model accepts text and image input and returns text output, so it handles both plain language prompts and prompts that include images alongside instructions. The context window is 262,144 tokens, which is large enough to hold long documents, extended conversations or sizeable code files in a single request. The maximum output is 235,929 tokens, so long generations are possible without truncating early. Supported request parameters include temperature, top_p, top_k, min_p, repetition_penalty, frequency_penalty, presence_penalty, max_tokens, stop, seed, logit_bias, logprobs and top_logprobs. It also supports response_format and structured_outputs for constrained output, plus tools, tool_choice and reasoning with include_reasoning for agent style and reasoning workflows.

Those capabilities point to a few practical uses. Long context plus image input suits document and screenshot analysis, where you paste a page or a UI capture and ask for an explanation or a summary. The visual coding angle mentioned by the provider makes it a reasonable candidate for reading screenshots of interfaces, diagrams or code and producing text descriptions or fixes. Tool and structured output support fits pipelines that need the model to call functions or return JSON shaped responses. Reasoning controls help on tasks where you want the model to work through a problem before answering, and the large output ceiling helps when the answer itself is long, such as a full file rewrite or a detailed report.

You call it through api14's OpenAI-compatible API using the model id moonshotai/kimi-k2.5, so existing OpenAI SDK code works with a base URL and key change. Pass the model id in the request, send messages with text or image content as needed, and set any of the supported parameters listed above. Because the API is OpenAI-compatible, parameters such as temperature, top_p, max_tokens, stop, seed, response_format, tools and tool_choice behave as you would expect, and the reasoning and structured output options are available when your client sends them.

Keep the limits in mind when choosing. Output is text only, so image generation or audio work needs a different model. If your task is short and simple, the large context and output ceilings are not advantages, and a smaller model may respond faster. The provider release date is 2026-01-27, so check that your tooling and expectations match that generation of the model.