Xiaomi MiMo-V2.6-Pro Takes Top Open-Weight AI Spot With Score of 46
Table of Contents
Xiaomi has launched its latest MiMo V2.6 AI model family, introducing new open-weight systems designed for advanced reasoning, multimodal tasks and AI agent workloads. The flagship MiMo-V2.6-Pro achieved a score of 46.32 on the Artificial Analysis Intelligence Index, placing it at the top of the open-weight models tracked by the index at the time of its release.
The new family includes MiMo-V2.6-Pro, MiMo-V2.6-Flash and MiMo-V2.6-Pro-UltraSpeed. Xiaomi has also released model weights, technical documentation and reinforcement-learning resources for developers and researchers.
MiMo-V2.6-Pro Reaches a Score of 46.32
The main highlight of the MiMo V2.6 AI release is the performance of the Pro model.
MiMo-V2.6-Pro recorded 46.32 on Artificial Analysis’ composite intelligence index. Xiaomi says this places it at the top of the open-weight category, ahead of other models such as Kimi K3 and GLM-5.3.
The result should be viewed within the context of the specific benchmark. It does not mean that MiMo-V2.6-Pro is ahead of every AI model available. Xiaomi itself notes that leading closed-source models remain ahead in some comparisons.
| Feature | MiMo-V2.6-Pro |
| Artificial Analysis score | 46.32 |
| Context window | 1 million tokens |
| Licence | MIT |
| Input price | $0.435 per million tokens |
| Output price | $0.87 per million tokens |
| Key capabilities | Multimodal reasoning and tool use |
The model supports multimodal inputs, long-context processing, reasoning and tool calling, according to Xiaomi.
MiMo-V2.6-Flash Targets Lower-Cost Workloads
Xiaomi has also introduced MiMo-V2.6-Flash, which is aimed at applications that need frequent AI requests at a lower cost.
Flash has approximately 309 billion total parameters, with around 15 billion active during inference. It retains a 1-million-token context window and is designed for large-scale professional workloads.
Xiaomi’s reported benchmark results show Flash performing competitively with Pro on several evaluations. However, these figures are company-reported results and should be distinguished from independent testing.
Flash is priced at $0.14 per million input tokens and $0.28 per million output tokens, making it considerably cheaper than the Pro model.
UltraSpeed Version Focuses on Faster Responses
The V2.6 family also includes MiMo-V2.6-Pro-UltraSpeed, a version aimed at applications where response time is particularly important.
Xiaomi says UltraSpeed can provide output speeds of up to 20 times faster than the standard Pro experience. The company positions it for real-time interaction and latency-sensitive workloads.
This gives developers different options within the same family: Pro for demanding reasoning tasks, Flash for lower-cost workloads and UltraSpeed for applications that prioritise response speed.
Large-Scale Reinforcement Learning
Xiaomi says reinforcement learning played a central role in developing the V2.6 models.
According to the company, the Pro and Flash models each completed 30 reinforcement-learning steps in less than six days, using approximately 750,000 trajectories. Xiaomi reported estimated training costs of around $2.62 million for Pro and $850,000 for Flash.
The company has also released more than 7,000 reinforcement-learning task environments, along with an end-to-end reinforcement-learning framework and additional development tools. These resources are intended to help researchers experiment with and build on the models.
Xiaomi Opens the MiMo V2.6 AI Family
The MiMo V2.6 AI release extends beyond model weights. Xiaomi has made the weights available through platforms including Hugging Face and released supporting technical resources.
MiMo-V2.6-Pro and Flash are released under an MIT licence, giving developers considerable flexibility to study and use the models within the terms of that licence.
The release could be useful for developers working on coding assistants, AI agents, research applications and multimodal systems.
However, benchmark results do not necessarily reflect performance in every real-world application. Independent testing will provide additional insight into how the models compare across different workloads.
What the MiMo V2.6 AI Release Means
Xiaomi’s latest release places the company more prominently in the open-weight AI market.
MiMo-V2.6-Pro’s Artificial Analysis score gives Xiaomi a notable benchmark result, while Flash and UltraSpeed broaden the family for developers with different cost and performance requirements.
At the same time, the model should not be described as the overall leader in AI. Its ranking is specific to the Artificial Analysis index and the open-weight category, and Xiaomi acknowledges that leading closed-source systems remain ahead in some comparisons.
Frequently Asked Questions
What is MiMo V2.6 AI?
MiMo V2.6 AI is Xiaomi’s latest family of AI models, consisting of MiMo-V2.6-Pro, MiMo-V2.6-Flash and MiMo-V2.6-Pro-UltraSpeed.
What score did MiMo-V2.6-Pro achieve?
It achieved 46.32 on the Artificial Analysis Intelligence Index, according to Xiaomi’s model documentation.
Is MiMo-V2.6-Pro open-weight?
Yes. Xiaomi has released the model weights under an MIT licence, along with supporting technical resources.
What is MiMo-V2.6-Flash?
Flash is the lower-cost model in the V2.6 family, designed for high-frequency and large-scale AI workloads.
What is MiMo-V2.6-Pro-UltraSpeed?
It is a faster version of the Pro model designed for latency-sensitive applications. Xiaomi says it can deliver up to 20 times the output speed of the standard Pro experience.
Conclusion
The MiMo V2.6 AI release represents Xiaomi’s latest move into advanced open-weight AI. MiMo-V2.6-Pro has reached 46.32 on the Artificial Analysis Intelligence Index, while Flash and UltraSpeed offer alternatives focused on cost and response speed.
With downloadable weights, an MIT licence and additional reinforcement-learning resources, Xiaomi is giving developers several ways to experiment with the new models. Their broader impact will become clearer as independent users evaluate them across real-world applications.
