“Impressive level of openness”: Xiaomi goes way beyond the usual open-weight playbook with MiMo-V2.6

The Chinese tech giant Xiaomi has unveiled MiMo-V2.6, an open-weight AI model that exceeds the typical open-weight playbook. This model boasts a staggering 1 trillion parameters, with 42 billion actively utilized at any given time, a context window of 1 million tokens, and the ability to process text, images, audio, and video. Xiaomi claims that MiMo-V2.6-Pro rivals the latest models from OpenAI and Anthropic, including GPT-6 Sol and Claude Opus 5.5, which have context windows of 1.05 million tokens and 1 million tokens, respectively.

Independent analysis corroborates Xiaomi's assertions, with Artificial Analysis assigning MiMo-V2.6-Pro an Intelligence Index score of 46, placing it at the top of the 114 large open-weight models it tracks.

The most intriguing aspect of Xiaomi's MiMo-V2.6 offering is the transparency and openness with which it trained the model. Xiaomi streamed its reinforcement learning (RL) training process in real-time on a public dashboard for five days, starting on September 15. The livestream displayed metrics from the production RL runs, revealing costs of $854,044 for the smaller MiMo-V2.6-Flash model and $2,620,670 for the Pro version. Despite these costs, the public RL bills are uncommon.

In addition to the model weights, Xiaomi has released the model under the permissive MIT license and made a 9-billion-parameter Qwen-based model available for further agentic RL research. Xiaomi has also shared a broader set of RL resources, including over 7,000 task environments covering software engineering, vulnerability reproduction, knowledge work, and web development.

A complete end-to-end training framework, covering environment interaction, reward evaluation, and policy optimization, has been made available. Lightweight agent harnesses for experimenting with various tools, prompts, and context setups are also included. However, Xiaomi has not yet publicly released the 7,000-plus environments and other supporting materials, stating that they will be shared "over the coming weeks."

The research community has responded positively to Xiaomi's unprecedented level of openness. Elie Bakouch, a former Hugging Face researcher now at Prime Intellect, lauded Xiaomi's commitment to releasing the 7,000-plus RL training data and the framework behind the top 6 model. Thomas Wolf, Hugging Face co-founder and chief science officer, emphasized that releasing many high-quality open-source RL environments is crucial for advancing open-source AI research, as they become increasingly vital in training models that rely on verifiable rewards (RLVR).