MiMo-V2.6-Flash

MiMo-V2.6-Flash is the fast, cost-efficient tier of Xiaomi's open-source MiMo V2.6 foundation family. Built on a Mixture-of-Experts architecture with 309B total parameters and 15B activated per token and a hybrid attention design, it delivers strong agentic, coding, and long-horizon performance at a fraction of the cost and latency of the flagship Pro tier. It is natively omnimodal, accepting text, image, video, and audio input and returning text, so interface screenshots and error states, architecture and network diagrams, dashboards and charts, recorded screen sessions, and audio can be reasoned over directly alongside source code and long textual context. A 1,048,576-token context window lets whole repositories, large document and log sets, long transcripts, or multi-session tool traces sit in a single call without a retrieval pipeline, while an output ceiling of up to 131,072 tokens supports long single-pass artifacts — complete reports, extended refactors and code generation, or detailed multi-step plans. It supports both thinking and non-thinking modes with selectable reasoning effort, native function calling, and structured (JSON) output, making it well suited to agentic tooling that plans, invokes tools, and returns machine-parseable results — coding assistants, computer-use and browser agents, and automated workflows that carry a lot of state across many turns. It reports 87.6 on Terminal-Bench 2.1, 80.8 on OSWorld-Verified, 73.6 on Toolathlon-Verified, and 67.9 on DeepSWE v1.1. As the Flash tier it is tuned for breadth and volume — bulk triage, large fan-out agent runs, and interactive assistants served at scale — where throughput and cost per token matter as much as raw peak intelligence, for which the heavier MiMo Pro tiers remain the better fit. It is available over the Chat Completions, OpenAI Responses, and Anthropic Messages APIs, so existing OpenAI- and Anthropic-style clients can target it without code changes.