Xiaomi has launched MiMo-V2.5, an open-weight multimodal model that treats audio, image and video as native tokens. It offers a 1,050,000-token context window and a pay-as-you-go price of $0.14 per million input tokens and $0.28 per million output tokens, targeting organisations that need on-premise, media-heavy AI.

Why a new multimodal approach matters

Most existing multimodal systems cheat. They first run a speech-to-text engine, then a vision model that turns pictures into captions, and finally feed the resulting text into a large language model (LLM). The LLM never sees the original sound wave or pixel data, so nuances such as a sigh, a pause, a facial twitch or a hand gesture disappear. MiMo-V2.5 skips the detour: it ingests audio and video directly as “tokens,” preserving timing, tone and visual cues that a transcript cannot capture.

The technical envelope

  • Token window: 1,050,000 tokens – enough to ingest multi-hour meeting recordings without chopping them into chunks.
  • Self-hosting: Deploy the model on a company’s own servers, so raw media never leaves the premises.
  • Pricing: $0.14 per million input tokens, $0.28 per million output tokens – a transparent cost that scales with usage.

In a quick sanity check of the text-only core, the model solved a classic merge-interval coding problem with the expected O(n log n) algorithm, computed a tank-fill rate math problem step-by-step, and extracted a JSON payload from an invoice without error. The audio and video pipelines were not exercised in that test.

Who stands to gain

Privacy-heavy industries like healthcare and finance cannot send raw audio or video to third-party APIs. By keeping the data behind the firewall, MiMo-V2.5 lets these organisations comply with data-protection rules while still gaining AI-driven insights. Potential applications include:

  • Meeting intelligence: Analyse entire video recordings to surface decisions, action items and speaker sentiment without a separate transcription stage.
  • Call-center analytics: Detect frustration, hesitation or confidence in a caller’s voice in real time.
  • On-device assistants: Build voice interfaces that understand tone and react to non-verbal cues without sending audio to the cloud.

The practical hurdles

MiMo-V2.5’s promise hinges on the performance of its audio and video encoders – components the author has not yet benchmarked. Enterprises must validate latency, accuracy and hardware consumption on their own datasets before committing to production. Teams that only need text will likely find a smaller, text-only model more efficient in both compute and cost. The open-weight nature of MiMo-V2.5 means the code and weights are publicly available; this enables deep customisation but also demands in-house expertise to maintain and secure the stack.

Takeaway

MiMo-V2.5 offers native media tokenisation, a huge context window and self-hosting at a clear per-token cost. For privacy-sensitive organisations that must keep raw audio and video in-house, it provides a viable pilot path. Real-world value will depend on how its audio and video encoders perform under enterprise workloads and whether the operational overhead of self-hosting fits within existing AI teams’ skill sets.