Xiaomi MiMo-V2.6 makes open-weight AI worth a production pilot
The highest-ranked open-weight model and sub-dollar output pricing change the build-or-wait calculation, but not the need for production evidence.
What launched
Xiaomi released MiMo-V2.6 Pro and Flash, including MIT-licensed weights. Pro accepts text, image, audio and video, offers a one-million-token context window, and is available through an API compatible with OpenAI and Anthropic protocols.
The production math
Artificial Analysis ranks MiMo-V2.6-Pro first among 114 open-weight models, with an Intelligence Index score of 46. At $0.435 per million input tokens and $0.87 per million output tokens, running a serious internal evaluation is now cheaper than waiting for consensus.
Our move
Virtual Arc would put MiMo behind a portable interface and route it a limited share of real production work. We would expand only where its cost per completed task remains lower after tool failures, retries and latency, while keeping an automatic fallback in place.
MiMo-V2.6 matters more to us than another closed flagship because it attacks both operating cost and supplier dependence. Pro ships with MIT-licensed weights, a one-million-token context window, and hosted API pricing of $0.435 per million input tokens and $0.87 per million output tokens. Virtual Arc would still not make it the default model on launch day. Cheap tokens are irrelevant if an agent misuses tools, loops, needs more retries or creates extra review work. We would place MiMo behind a provider-neutral routing layer, send it a controlled share of real tasks and compare cost per successfully completed job after retries, latency and intervention are included. We would retain an established fallback and consider self-hosting only when data-control requirements or sustained volume justify the substantial infrastructure demanded by a model of this size. This release is worth touching now, but as a measured production pilot rather than a migration.