Global technology company Xiaomi Corp. today announced the release and open-sourcing of its MiMo-V2.6 series generative artificial intelligence models today, which includes two natively omnimodal models bringing a balance of intelligence, efficiency and cost.
The MiMo-V2.6 series includes two natively omnimodal models with a new flagship model, MiMo-V2.6-Pro, and a smaller, efficient Flash variant. The company also said it’s rolling out a slimmed-down, Pro-UltraSpeed variant that delivers up to 20 times faster output than Pro at the same quality for extreme generation speeds.
The company described the Pro and Flash models as native “omnimodal”: accepting text, image, video and audio in the same family, with 1 million-token context windows. Pro’s published architecture lists a 681 million-parameter vision encoder including a 308 million-parameter AudioTokenizer and 127 million-parameter audio patch encoder capable of distinguishing speech.
This design enables the larger models to meet today’s growing needs for agentic AI use cases and task instructions, particularly for computer use. For example, a computer-use agent could read a task instruction, read a screenshot of a user interface or video, reason about it, then decide what to do next within one model loop.
A coding or design agent can combine source code and a lengthy repository context with screenshots, UI mockups and reference images all in the same session. For business use, the model could ingest numerous documents, including multimodal components such as call audio, screenshots, logs and notes, without requiring tools to break down transcripts or orchestrate each modality.
In the announcement blog post, the company demonstrated how well the model managed building game worlds, Blender-based 3D modeling, embodied simulation and video-music workflows.
The company said V2.6-Pro scored 46.32 on the Artificial Analysis Intelligence Index, ahead of Kimi K3 and Qwen3.8 Max, the highest-ranked open-source/open-weight model in that comparison at launch. However, the model still trails the most advanced flagship proprietary closed-source models on the market, Claude Fable 5.1 and GPT-6 Astra, on aggregate measures.
On agentic tasks, V2.6-Pro showed strong results in testing, close to or occasionally above leading closed systems: 53.1 on AutomationBench versus Claude Opus 5’s 50.3; 31.6 on Agents’ Last Exam, tying Opus 5; and 89.9 on Terminal Bench 2.1, slightly ahead of Opus 5’s 89.1.
Cost to efficiency benefit
Xiaomi said the V2.6 series will retain the standard application programming interface pricing used for MiMo-V2.5, and the model cards also list availability through the company’s AI Studio, MiMo Desktop and MiMo Code.
The Pro and Flash models are also available on OpenRouter with 1.05-million-token context through the unified API.
This means V2.6-Flash is available at $0.14 per 1 million input and $0.28 per output and Pro for $0.435 and $0.87. Whereas Pro-UltraSpeed kicks all the way up 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, taking into account cached-token costs, which are significantly reduced. This is versus $10 to $50 input and output per million tokens for OpenAI Group PBC GPT-6 Astra and Anthropic PBC’s Claude Fable 5.1. For cache-heavy agents, that gap widens further, although the relevant business comparison is less about price per token and more about price per task, determined by reliability and how many loops it took to complete successfully.
