Xiaomi Releases MiMo-V2.6 Open-Weight Models

Xiaomi released MiMo-V2.6 Pro and Flash, with the Pro version reaching 46 points on the Artificial Analysis Intelligence Index. This achievement places it at the top of open-weight models, marking a substantial improvement over the previous V2.5 iteration.

The models are native multimodal, handling text, image, audio, and video inputs. Beyond the models, Xiaomi provided training code, reinforcement learning environments, and evaluation frameworks to support reproducibility and research expansion.

Performance gains were driven by large-scale reinforcement learning over six days. Xiaomi reported training costs of $2.62 million for Pro and $0.85 million for Flash, resulting in significant score increases across DeepSWE benchmarks.

MiMo-V2.6 Pro tied with Grok 4.7 at 46 points on the Artificial Analysis index. It remains behind proprietary models like GPT-5.6 Sol, Opus 5, and GPT-6 Astra in this specific metric.

The Pro model excels in coding and tool-use benchmarks, though it trails slightly behind GPT-5.6 Sol and Claude Opus 5 in specific tests like DeepSWE. It performed strongly in knowledge-based tasks, surpassing several major competitors.

Xiaomi's API pricing is significantly lower than major competitors like Grok 4.7 or Claude Fable 5.1. While token costs are low, actual operational expenses may vary based on specific task requirements and model usage patterns.

Artificial Analysis estimated the cost per task for MiMo-V2.6 Pro at $0.13. The model demonstrated an output speed of 134.3 tokens per second with a time-to-first-token of 2.15 seconds.

The models support 3D spatial reasoning and computer control. Demonstrations include 3D modeling in Blender, design adjustments in Canva, and spreadsheet manipulation, showcasing broad software interaction capabilities.

MiMo-V2.6 Pro reached 8th place overall in the Design Arena, ranking 3rd among open-weight models. It showed marked improvements in website creation and agentic front-end development tasks.

Xiaomi highlighted the model's role in identifying MOF materials for PFAS adsorption and formalizing complex mathematical proofs in Lean 4. The model successfully passed kernel verification for its mathematical work.

The model utilizes a Mixture-of-Experts (MoE) architecture with 1.02 trillion total parameters, activating 42 billion per inference. While efficient, it still requires significant infrastructure for self-hosting.

Detailed documentation and technical specifications regarding the MiMo-V2.6 release are available on the official Xiaomi project website.