Xiaomi released MiMo-V2.6-Pro, an open-weights omnimodal model, alongside MiMo-V2.6-Flash and an UltraSpeed variant. The model was reportedly trained for about $3M and is described as the new top open-weights model, with Xiaomi scaling RL compute across larger batches and richer multi-task environments.
SynthePulse Insight · AI deep readingMembers
Version 1 · 1 source
Xiaomi releases the natively omni-modal open-weight MiMo-V2.6 series; Artificial Analysis says its Pro version tops the open-weight intelligence index with 46 points, and with figures like roughly $3 million in training cost and a 130-hour RL run, it pushes "post-training/RL as a cheaper path to the frontier" onto the table.
The hardest fact this round comes from Artificial Analysis's evaluation basis: MiMo-V2.6-Pro tops its intelligence index for open-weight models with 46 points, at 1.02T total/42B active parameters, priced at $0.435/M input and $0.87/M output, and sits on the "intelligence vs cost per task" Pareto frontier at $0.13 per intelligence-index task. The organization also says this is Xiaomi's first time in this position. What needs to be distinguished is that this is the evaluator's basis and conclusion, not an independent third-party reproduction.
Xiaomi's own statement is that the MiMo-V2.6 series contains two natively omni-modal models: Pro is its strongest to date, and Flash strikes the best balance among intelligence, efficiency and cost; it also launches MiMo-V2.6-Pro-UltraSpeed, claiming up to 20x faster output at the same quality. According to a paraphrase, the model uses the MIT license. These are all statements by the releaser or the paraphraser.
A notable structural backdrop is that Xiaomi is traditionally not regarded as one of the "Chinese AI six tigers," so its taking the top spot is a surprise to the established roster order; at the same time, after former DeepSeek engineer Fuli Luo joined Xiaomi, the final RL training run was livestreamed, described as unusually high internal-metric transparency. This shows the competitive dimension is expanding from "the model itself" to "process visibility."
This article strictly follows the given sources. Artificial Analysis's index score, parameters and pricing are the evaluator's basis; Xiaomi's model positioning, open-source scope and UltraSpeed acceleration are the releaser's statements; the $3M, 130 hours/75B tokens/$2.6M, MIT license, 7k+ environments and so on come from paraphrases or headlines, have not obtained first-party or independent verification, and are all handled as source_claim.
Primary report
Primary source