08:49Machine-translated from Azerbaijani by AIRead in Azerbaijani
Xiaomi unveils MiMo-V2.6-Pro AI model with 1 trillion parameters
Xiaomi has unveiled its new MiMo-V2.6-Pro large language model. The company’s new AI model has 1.02 trillion parameters, a 1 million-token context window and the ability to process text, images, audio and video. The model is particularly notable for its capabilities designed for programming and AI agents.
Xiaomi says MiMo-V2.6-Pro uses a Mixture of Experts (MoE) architecture. As a result, not all of the model’s 1.02 trillion parameters are activated for every request. Approximately 42 billion parameters are activated during inference. This approach reduces demand for computing resources while preserving the capabilities of the large model.
The model also has multimodal processing capabilities. In addition to text, MiMo-V2.6-Pro can accept and process images, audio and video within the same system. It uses Xiaomi’s 681 million-parameter MiMo-ViT encoder to process images and a separate audio encoder to process audio data.
One of the new model’s main features is its 1 million-token context window. This allows large documents, extensive codebases and long-running AI agent tasks to be processed within the same context.
Xiaomi placed particular emphasis on reinforcement learning (RL) when developing the model. With this method, the model does more than learn from pre-prepared examples; it improves its behavior based on the results it obtains while completing tasks. This approach is particularly important for AI agents that need to plan, use various tools, correct errors and execute long-running tasks.
According to the performance tests cited by the source, MiMo-V2.6-Pro can compete with the current leading open models on some programming tasks and outperforms DeepSeek’s latest model in certain tests. However, user feedback indicates that results vary depending on the task, and it is still too early to say that MiMo-V2.6-Pro is more powerful in every area.
Xiaomi has also introduced the MiMo-V2.6-Flash model, which focuses on lower computing costs. The Flash version has 310 billion parameters in total, with approximately 15 billion parameters activated during inference. This model also supports a 1 million-token context window and text, image, audio and video inputs.
According to Xiaomi’s published API pricing, 1 million input tokens for MiMo-V2.6-Flash cost $0.14, while output tokens cost $0.28. For MiMo-V2.6-Pro, the corresponding prices are $0.435 and $0.87, respectively. The Flash version is therefore positioned as a lower-cost alternative for high-volume programming, document processing and AI agent systems.