MiMo-V2.6-Pro: New model leads open-weight ranking

Xiaomi has released a new generation of its open AI models with MiMo-V2.6. The focus is on the top model MiMo-V2.6-Pro, which takes the lead among open-weight models in Artificial Analysis’s Intelligence Index. It now ranks ahead of Chinese competitors such as GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, and reaches the level of some powerful proprietary models, such as xAI’s newly released AI model Grok 4.7.

At the same time, despite its high benchmark performance, the model remains comparatively inexpensive: Xiaomi charges $0.435 per million input tokens without cache hits and $0.87 per million output tokens via its API. With a cache hit, the input price drops to $0.0036.

In the measurements by Artificial Analysis, the favorable price-performance ratio is particularly evident: a task in the Intelligence Index costs an average of only $0.13 with MiMo-V2.6-Pro, significantly less than with similarly powerful open-weight and proprietary models.

With MiMo-V2.6-Flash, Xiaomi is also releasing a smaller and even cheaper variant of the new model family. Via the API, it costs $0.14 per million input tokens without cache hits and $0.28 per million output tokens. With a cache hit, the input price is $0.0028.

According to Xiaomi, the Flash variant achieves values close to the Pro model in several agent benchmarks. An independent evaluation by Artificial Analysis is not yet available.

Highly Scaled Reinforcement Learning

Technically, Pro and Flash differ mainly in model size. MiMo-V2.6-Pro is based on a Mixture-of-Experts architecture with a total of 1.02 trillion parameters, of which 42 billion are activated per token. Flash has a total of 309 billion parameters and 15 billion active parameters. Both models process text, images, video, and audio, support up to one million tokens of context, and are available with open weights.

According to Xiaomi’s own benchmarks, the two new models show significant progress compared to MiMo-V2.5-Pro, released in April, in several coding, agent, and cybersecurity tests. The company sees a key reason in downstream Reinforcement Learning (RL): Pro and Flash repeatedly processed complex tasks from these areas, completing 30 RL steps each with around 750,000 task runs, partly depicting long, multi-stage agent processes. Xiaomi describes the RL method in detail in the technical report. In addition to the model weights, Xiaomi also provides the RL code and the training environments used for further research.

With MiMo-V2.6-Distill-Qwen-9B, Xiaomi is also releasing a small research model based on Qwen3.5-9B. The model is intended as a starting point for research into how agent capabilities can be transferred to smaller models and subsequently expanded with RL.

MiMo-V2.6-Pro and MiMo-V2.6-Flash are available via Xiaomi’s API. Xiaomi also provides the weights of all three models under the MIT license on Hugging Face.

(wpl)