Xiaomi has released and open-sourced its MiMo-V2.6 series of generative AI models: two natively omnimodal models, MiMo-V2.6-Pro and a smaller Flash variant, plus a Pro-UltraSpeed build that delivers up to 20 times faster output than Pro at the same quality. The Pro model scores 46.32 on the Artificial Analysis Intelligence Index, making it the highest-ranked open-weight model at launch. This matters because Xiaomi is now competing directly with closed flagship systems on agentic tasks while keeping the weights open.
What the series includes and how it is priced
Besides Pro and Flash, Xiaomi is rolling out Pro-UltraSpeed, a slimmed-down variant that trades cost for generation speed. Pricing stays at the standard MiMo-V2.5 API rates: Flash costs $0.14 per 1 million input tokens and $0.28 per output token, Pro costs $0.435 and $0.87, and Pro-UltraSpeed jumps to $4.35 and $8.70 for the 20-times faster generation. Xiaomi claims Pro costs roughly one-20th to one-60th as much as overseas models at comparable intelligence once cached-token costs are counted, against $10 to $50 per million tokens for GPT-6 Astra and Claude Fable 5.1. The models are available through Xiaomi's AI Studio, MiMo Desktop and MiMo Code, and both Pro and Flash are on OpenRouter with a 1.05-million-token context via a unified API.
Both Pro and Flash are described as natively omnimodal: they accept text, image, video and audio in the same model family and carry 1 million-token context windows. The published Pro architecture lists a 681 million-parameter vision encoder, which includes a 308 million-parameter AudioTokenizer and a 127 million-parameter audio patch encoder capable of distinguishing speech. Keeping these components inside one model removes the need for separate transcription or modality-orchestration tools. In practice, a business process that mixes call audio, screenshots, logs and notes can be handled in a single session rather than split across a pipeline.
The design targets agentic workloads directly. A computer-use agent can read a task instruction, take in a screenshot or video of an interface, reason about it and decide the next step within one model loop. A coding or design agent can hold source code, a long repository context, screenshots, UI mockups and reference images in the same session. Xiaomi's announcement blog demonstrated game-world building, Blender-based 3D modeling, embodied simulation and video-music workflows. On benchmarks, V2.6-Pro scored 53.1 on AutomationBench against Claude Opus 5's 50.3, tied Opus 5 at 31.6 on Agents' Last Exam, and reached 89.9 on Terminal Bench 2.1 versus 89.1 for Opus 5.
What this means for companies adopting AI
For teams building agents, the practical change is the price of a completed task rather than the price of a token. Xiaomi's own framing is that the relevant comparison depends on reliability and how many loops an agent needs to finish successfully, and cache-heavy workloads widen the gap further because cached tokens are priced significantly lower. A small company can now run multimodal agent workloads on open weights at rates that were previously reserved for much simpler models, while a large organization can compare Pro-UltraSpeed against closed systems when turnaround time matters more than unit cost.
Several points still need verification before a vendor decision. The 20-times speed claim applies to Pro-UltraSpeed at the same stated quality, but that variant costs ten times more per token than Pro, so the economics depend on the actual task mix. The headline index score of 46.32 still trails Claude Fable 5.1 and GPT-6 Astra on aggregate measures, which matters for workloads that mix agentic and general reasoning. OpenRouter exposes a 1.05-million-token context while Xiaomi lists 1 million tokens, so the effective window should be confirmed per channel. Reliability across long sessions, not the per-token list price, remains the number to watch.
The clearest marker to follow is whether Pro-UltraSpeed adoption shows up in real agent deployments rather than in benchmark runs. If teams report completing computer-use and coding tasks in fewer loops at the stated $4.35 and $8.70 per million tokens, the open-weight price advantage will hold under production load; if not, the comparison shifts back to closed flagships regardless of list price.
