Xiaomi has released the MiMo-V2.6 series, introducing natively omnimodal models including MiMo-V2.6-Pro and MiMo-V2.6-Flash. The company also launched MiMo-V2.6-Pro-UltraSpeed, which delivers up to 20x faster output speed at the same quality.
- MiMo-V2.6-Pro is described as Xiaomi's most capable model to date, debuting as the top open-weights model on Artificial Analysis's Intelligence Index with a score of 46.
- The model features 1.02T total parameters with 42B active, trained using RL compute scaled along larger batches, richer environments, and more grader compute.
- Xiaomi published its final RL training runs live and is open-sourcing the tooling, including environment code and training recipes for coding, cyber, general, visual, and music tasks.
- The models are available under the MIT license, with input costs at $0.435/M tokens and output costs at $0.87/M tokens.
The release highlights Xiaomi's entry into the frontier AI lab space and its commitment to transparency through open-sourced training environments and detailed technical reports.