Overview
MiMo-V2.5-Omni is Xiaomi's **native omnimodal perception model** in the MiMo-V2.5 family. It takes over from MiMo-V2-Omni, retired on June 30, 2026, and now serves as the primary entry point for Xiaomi's omnimodal capabilities. "Native" means images, video, audio, and text are jointly modeled at the architectural level, so cross-modal reasoning is a default capability rather than a set of encoders bolted onto a text model.
On specs, MiMo-V2.5-Omni supports a **1M context window with up to 128K output tokens**, at rate limits of RPM 100 / TPM 10M. It suits workloads you can feed in one pass: tens-of-minutes-long videos, long meeting recordings, and full illustrated manuals.
Together with the same series' **MiMo-V2.5-Pro** (reasoning/coding flagship), **MiMo-V2.5-Flash** (cost-efficient lightweight tier), **MiMo-V2.5-TTS** (speech synthesis with Base TTS / VoiceDesign / VoiceClone sub-models), and **MiMo-V2.5-ASR** (speech recognition with solid dialect coverage), it forms Xiaomi's complete multimodal matrix. The whole V2.5 series ships under the **MIT license** and offers both OpenAI- and Anthropic-compatible endpoints, keeping migration costs low.
Key Features
- Native Omnimodal Joint Modeling: Image/video/audio/text understood jointly at the architectural level, with cross-modal reasoning as a default capability rather than an add-on module
- 1M Context Window: Supports 1M-token context with up to 128K output tokens, handling long videos, long audio, and multi-image documents in a single pass
- Video and Long Audio Understanding: Native support for understanding, summarizing, Q&A, and key-segment localization on tens-of-minutes-long videos and audio
- Interleaved Image-Text Reasoning: Handles interleaved input of "text + image + text + image" for complex reasoning, covering papers, technical documents, and multi-image manuals
- Dual-Protocol Compatible Endpoints: Offers both OpenAI- and Anthropic-compatible interfaces, so existing codebases can switch over with almost no code changes
- Open Source under MIT License: V2.5 series weights and inference code are open under the MIT license, allowing commercial use and fine-tuning
- End-Cloud Synergy Deployment: A distilled version can run on Xiaomi HyperOS on-device, while the cloud version provides full capabilities, supporting a consistent experience across Xiaomi phones, cars, and smart home devices
Use Cases
- Full-modal assistant capabilities on Xiaomi phones/cars/smart home devices
- Video content understanding: automatic summarization, chapter segmentation, Q&A, keyframe localization
- Transcription, segmentation, and key-point extraction for long meeting recordings
- Education scenarios: interleaved reasoning with images, text, and formulas, subject Q&A, video lecture explanations
- Industry/manufacturing: defect recognition directly from camera video or images, plus text report generation
- Accessibility: image description, video commentary, cross-modal dialogue, benefiting visually/hearing-impaired users
- Developers fine-tuning industry-specific full-modal models based on open-source weights
Pros
- Native omnimodal architecture, with cross-modal capabilities among the top tier of domestic models
- 1M context with 128K output gives ample capacity for long video and long audio scenarios
- MIT license open-sourcing is one of the more open release stances among major domestic companies
- OpenAI- and Anthropic-compatible endpoints keep integration and migration costs low
- Combined with V2.5-Pro/Flash/TTS/ASR, covers reasoning, coding, omnimodal, and speech scenarios comprehensively
- End-cloud synergy: on-device distillation plus full cloud version, amplifying value across Xiaomi's entire product matrix
Pricing
**Open-source weights are free** (MIT license, downloadable from HuggingFace/GitHub). **Official API** (mimo.xiaomi.com) offers a free trial quota, with overage billed per token and rate limits of RPM 100 / TPM 10M. Xiaomi phone/car users can use full-modal capabilities **completely free** in system-integrated scenarios. Enterprise private deployment can be quoted per GPU node.
Summary
MiMo-V2.5-Omni is one of the more solidly specified domestic open-source omnimodal models available—**native unified architecture + 1M context + MIT open source + dual-protocol endpoints**. If you are building cross-modal applications (video Q&A, multimodal agents, comprehensive image-text-audio understanding), Omni is the open-source choice you can pick up today. For reasoning and coding, consider the same series' **MiMo-V2.5-Pro**; for speech synthesis and recognition, use **MiMo-V2.5-TTS** and **MiMo-V2.5-ASR** respectively.
Version History
- MiMo-V2.5-Omni Becomes the Omnimodal Mainstay (2026-06-30): The MiMo-V2 series (V2-Pro / V2-Omni / V2-Flash) was officially retired at 00:00 Beijing time on June 30, 2026, and the original model names stopped working. Xiaomi had already been routing V2 series requests to the V2.5 series and billing at V2.5 pricing, making V2.5 the only in-service generation, spanning Pro / Flash / Omni / TTS / ASR.
- MiMo-V2.5 Series Released (2026-05-20): Xiaomi released the full MiMo-V2.5 lineup: V2.5-Pro reasoning/coding flagship, V2.5-Flash cost-efficient tier, V2.5-Omni native omnimodal (1M context / 128K output / RPM 100 / TPM 10M), V2.5-TTS with three sub-models (Base TTS / VoiceDesign / VoiceClone), and V2.5-ASR open-source dialect recognition. All under the MIT license with both OpenAI- and Anthropic-compatible endpoints.