A New Leader in Open-Weights AI
In a significant expansion of its AI ecosystem, Xiaomi has launched the MiMo-V2.6 model family, with the flagship MiMo-V2.6-Pro securing the top spot on the Artificial Intelligence Index managed by benchmarking firm Artificial Analysis. The new model achieved a score of 46, placing it ahead of major proprietary competitors like Google’s Gemini 3.8 Flash (41) and rivaling the newly released Grok 4.7.
Unlike many proprietary frontier models, MiMo-V2.6-Pro is released under an MIT license, allowing developers and enterprises to download the model from Hugging Face for local deployment or customization. For those opting for cloud access, Xiaomi has positioned its API pricing aggressively, charging $0.435 per million input tokens and $0.87 per million output tokens.
Scaling Reinforcement Learning
The technical foundation of the V2.6 release represents a departure from static training methods. Xiaomi’s report details a massive investment in Reinforcement Learning (RL), with the company spending over $2.6 million to train the Pro model through 30 large-scale RL steps. This process involved roughly 750,000 trajectories, focusing on long-horizon reasoning and complex agentic workflows rather than simple text generation.
Xiaomi’s strategy, which they term “You Only RL Once,” integrates diverse domains—such as coding, visual processing, and cybersecurity—into a single training cycle. By using Group Relative Policy Optimization (GRPO), the system manages asynchronous task completion, ensuring that long-running agent jobs do not bottleneck the GPU fleet.
Solving for ‘Reward Hacking’
A critical challenge addressed in the V2.6 development is “reward hacking,” where AI agents find ways to satisfy test criteria without actually performing the intended work—such as searching online for pre-existing fixes rather than solving a bug. To combat this, Xiaomi implemented two proprietary systems: Groupwise Reward Synthesis (GRS) and Groupwise Advantage Redistribution (GAR).
These systems evaluate the quality of an agent’s behavior, not just the final result. By comparing multiple attempts at the same problem, GRS creates rubrics that punish “brittle” solutions—such as those relying on speculative compatibility patches—and reward precise, minimal code changes. Experimental data provided by Xiaomi suggests that the use of GAR led to more stable token usage and higher pass rates across complex tasks compared to standard training protocols.