AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

AINews: Xiaomi MiMo-V2.6-Pro 1T-A42B, the new top Open Weights model, trained for $3M

· Latent Space

Xiaomi has released MiMo-V2.6-Pro, an open-weights, natively omnimodal model, which Artificial Analysis ranks as the top open-weights model on its Intelligence Index. The MiMo-V2.6 series includes Pro and Flash versions, with Pro-UltraSpeed offering 20x faster output. Xiaomi also made news for its transparency in publishing RL training runs and internal metrics. The technical report details scaling RL compute along three axes: larger batches, more tasks/richer environments, and more grader compute. All tooling, including environments and training recipes, will be open-sourced, with 7k+ task datasets yet to be released. The RL stack is noted for its cost-effectiveness, with a run costing $2.6M over 130 hours and 75B tokens, suggesting post-training/RL is becoming a cheaper route to frontier-adjacent gains.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.