Xiaomi: MiMo-V2.5

Xiaomi: MiMo-V2.5 is a text model from Xiaomi, listed on api14 under the model id xiaomi/mimo-v2.5. Xiaomi describes it as a native omnimodal model that delivers Pro-level agentic performance at roughly half the inference cost, and that surpasses MiMo-V2-Omni in multimodal perception across image and video understanding. It sits in the text category, so text is what you get back, but the range of inputs it accepts is much wider than the category label suggests.

On the input side, MiMo-V2.5 accepts text, audio, image and video, and it produces text output. That makes it a model you can point at mixed media rather than only at written prompts. The context window is 1,050,000 tokens, which is large enough to hold very long documents, long transcripts or extended media-derived context in a single request. The maximum output is 131,072 tokens, so long generations are possible within the same call. Reasoning is part of the picture too: the supported parameters include reasoning and include_reasoning, so you can work with the model's reasoning behaviour rather than treating it as a plain text generator.

Those capabilities point to a fairly specific set of use cases. Because audio, image and video are all accepted as input, MiMo-V2.5 fits tasks such as summarising or querying recorded meetings and other audio, extracting information from images, and answering questions about video content, all without a separate transcription or vision step in front of the model. The very large context window suits work over long source material, for example analysing a lengthy report alongside supporting media, or keeping a long agent session in view. Since the model is described as agentic, tool use is a natural fit as well, and the parameter list backs that up.

You call it through api14's OpenAI-compatible API using the model id xiaomi/mimo-v2.5. The request parameters it supports are frequency_penalty, include_reasoning, max_tokens, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k and top_p. In practice that means you can control sampling with temperature, top_p and top_k, shape output with response_format or structured_outputs, cap generation with max_tokens, make runs reproducible with seed, and define function calling through tools and tool_choice. Reasoning can be requested or surfaced via the reasoning and include_reasoning parameters.

The main thing to keep in mind is the output side. MiMo-V2.5 returns text only, so if your application needs generated images, audio or video, this model will not produce them, even though it can read those formats as input. Its category is text, so it is best treated as a multimodal-input, text-output model. For workloads that are purely text and do not need audio, image or video understanding, a simpler text model may be a more direct fit, and the large context window and omnimodal input handling are the features worth weighing when you decide.