MiMo-V2.5: Open-Weight Omnimodal AI With Audio and Vision

MiMo-V2.5 is an open-weight Mixture-of-Experts model from Xiaomi for agentic work and multimodal understanding. It runs 15B active parameters out of 310B, takes a 1M-token context, and is natively omnimodal: text, image, video, and audio in, text out.

It's the only open-weight model in the is*ai lineup that reads audio, and its focus is sharp perception (charts, documents, video) alongside frontier-level agentic ability.

What Makes MiMo-V2.5 Different

• 310B total parameters, 15B active. Sparse MoE tuned for token efficiency, so you get omnimodal, agentic capability at a lower cost per token than a frontier-scale flagship.
• Native omnimodality, including audio. A 729M vision encoder and a dedicated audio encoder give it image, video, and audio understanding in one architecture, with audio handled by a purpose-built encoder rather than tacked on.
• 1M-token context. Room for full documents, long transcripts, or extended agent sessions in one pass, on a hybrid sliding-window attention backbone from MiMo-V2-Flash.
• Strong multimodal perception. Xiaomi reports it stays level with frontier closed-source models on image, video, chart, and document understanding.
• Frontier agentic capability. Post-trained with supervised fine-tuning, large-scale agentic RL, and multi-teacher distillation for tool use and multi-step tasks.
• MIT license. Trained on roughly 48T tokens in FP8, with weights you can download and use commercially without authorization.

Where MiMo-V2.5 Delivers

With is*smart, MiMo-V2.5 is already deployed and optimized inside is*hosting, so the hardware a 310B multimodal model needs is handled. Here is where it earns its place.

Audio and Cross-Format Understanding

The audio encoder is what sets it apart. It can work over meeting recordings, calls, and voice notes alongside text, images, and video, which fits tasks that mix formats: summarizing a recorded call against a slide deck, or answering questions across a document and its audio together.

Visual Reasoning and Document Analysis

MiMo-V2.5 reads charts, tables, scanned documents, and screenshots and reasons over what it sees. That makes it a fit for visual question answering, chart and figure analysis, and pulling structured meaning out of documents that aren't plain text.

Agentic Work That Needs to See

Many agent tasks stall because the model can't interpret what's on the screen. MiMo-V2.5 pairs perception with tool use, so an agent can read a dashboard, interpret a chart, and act on it in the same loop instead of handing off to a separate vision step.

Cost-Efficient Multimodal at Scale

Because it's efficient for its size, perception-heavy pipelines that would otherwise need a frontier-priced model can run here at a lower cost per task, without dropping to a model that can't see or hear.

When Another Model Fits Better

MiMo-V2.5 is built around perception. When you don't need it, a leaner model is a better match:

• Pure text coding and agents, no images, audio, or video? DeepSeek V4 Flash is cheaper and focused on exactly that.
• The heaviest, longest frontier coding and agentic runs with image and video? MiniMax M3 is the heavier option.
• High-volume lightweight tasks like translation, moderation, and classification? Gemini 3.1 Flash-Lite is priced for that scale.

A Few Practical Notes

MiMo-V2.5 reads image, video, and audio but outputs text only, so it understands media rather than generating it. It's a large model, so perception-heavy calls cost more than a Flash-tier text model. For plain text work, a lighter model is cheaper. Match it to tasks that actually use its perception or audio and it earns its keep; otherwise you're paying for capability you won't touch.

MiMo-V2.5 and is*smart

MiMo-V2.5 is MIT-licensed, so the weights are free to run and build on. Through is*smart, it's already hosted and optimized inside the is*hosting infrastructure, so your files, prompts, and media stay in the environment: no third-party APIs, no data leaving, and no vendor lock-in. You get an omnimodal, agentic model without sourcing the GPUs a 310B model needs.

Subscribe to is*smart to get instant access to MiMo-V2.5 and put audio, vision, and agentic reasoning behind a single model.