Xiaomi's MiMo-V2.6-Pro leads open-weights models at low cost per task

Xiaomi's MiMo-V2.6-Pro is an open-weights model, meaning anyone can download and run it, and Artificial Analysis scored it at 46 on its Intelligence Index when Xiaomi released it in September. That made it the top open-weights model on the index and a large step up from its predecessor, MiMo-V2.5-Pro, which scored 26. Xiaomi kept the predecessor's prices: $0.435 per million input tokens, with a 99% discount on cache hits, and $0.87 per million output tokens.

The stronger claim is about cost. Running Artificial Analysis's index cost $0.13 per task, and the firm says that places the model on its Pareto frontier for intelligence against cost per task, the set of models that no other tested model beats on score and cost at once. Cost per task combines the per-token prices with how many input, cached and output tokens a model actually spends on each task, as Artificial Analysis describes on its evaluation pages, so it is a closer guide to the bill for real work than the price list alone. The model is a mixture of experts with 1.02 trillion parameters in total, of which 42 billion are active, meaning only that fraction of the network is used for each token it processes.

The open-weights lead was provisional. In an October 23 article, Artificial Analysis said Moonshot AI's Kimi K3.1 Pro, which scored 55, will become the highest-scoring open-weights model once its weights are published, narrowing the gap between the best open-weights and best proprietary scores on the index from 12 points to 3. On October 27 the firm moved to version 5 of the index, adding two benchmarks and upgrading three, so MiMo-V2.6-Pro's 46 belongs to the earlier version and measures a different set of tasks from v5 scores. What holds is the narrower finding: at release, a downloadable model reached the top of the open-weights field at a very low cost per task, without a price increase over a much weaker predecessor.