Xiaomi Livestreams the Training Process of Its MiMo-V2.6 Model with Publicly Disclosed Costs and Steps

Xiaomi revealed its new MiMo-V2.6 model based on reinforcement learning technology, choosing an unusual presentation style by livestreaming the training process with public disclosure of costs and the number of training steps, in what appears to be a marketing move aimed at the research and developer community.

With this model, Xiaomi focuses on reasoning capabilities enhanced by reinforcement learning (RL), a technology that has gained remarkable traction in recent months following the success of the Chinese DeepSeek-R1 model. The announced figures point to mid-range training data and model sizes; however, the model has not yet been released for public use or independent testing.

It is worth noting that publicly disclosed training figures provide limited informational value without independent comparison against standardized benchmarks, as the gap between training costs and actual performance in real-world applications remains large and cannot be estimated from raw numbers alone. In the absence of independent evaluation results, it is difficult to determine this model's position in the competitive landscape.

This launch falls within the context of escalating Chinese competition in the field of AI models, engaged in by companies such as Alibaba, Baidu, Tencent, and ByteDance, alongside a wave of startups racing to fill the gap created by the U.S. decision to restrict advanced chip exports. The results of independent evaluations following the official release will determine the true value of this model.

**What do these terms mean?**

- **Reinforcement Learning (RL)**: A training method where the model is rewarded for correct answers and penalized for incorrect ones, improving its problem-solving and reasoning capabilities.

- **Benchmarks**: Standardized tests that measure a model's capabilities in areas such as mathematics, programming, and language comprehension, used to objectively compare different models.