Goldman Sachs has released a report stating that Xiaomi's MiMo recently conducted a live stream showcasing the reinforcement learning (RL) training process of MiMo-V2.6. Remarkably, the V2.6-Flash and V2.6-Pro variants ceased training after just 30 steps, yet achieved impressive top scores of 67.86 and 72.57, respectively, in the DeepSWE benchmark test. This RL training substantially boosted the performance of long-horizon agents. Goldman Sachs notes that this marks a pioneering moment, as it is the first instance a prominent AI lab has openly shared post-training telemetry data in real-time. It is anticipated that pricing will emerge as a key consideration upon the official launch of MiMo-V2.6. This iteration is poised to reinforce Xiaomi's standing in agent coding and multimodal integration, offering greater potential for API pricing strategies. Xiaomi could potentially shake up the prevailing market pricing framework, following in the footsteps of DeepSeek. Furthermore, Goldman Sachs predicts that MiMo-V3, which will incorporate the innovative HySparse architecture, is slated for release in early 2027. The sparsity ratio of this new model will escalate from 7:1 in V2/V2.5 to 11:1, thereby enhancing efficiency and driving down inference costs. Goldman Sachs reiterates its 'Buy' stance on Xiaomi, setting a 12-month target price of HK$39.