Xiaomi debuted MiMo-V2.6, a trillion-parameter AI model that the Chinese tech company trained while broadcasting its reinforcement-learning costs in real time, according to a report published this week by The New Stack. The flagship MiMo-V2.6-Pro features 42 billion active parameters, handles a one-million-token context window, and works with text, images, audio and video. Independent tracking platform Artificial Analysis ranked it first among 114 large open-weight models, awarding it an Intelligence Index score of 46.
Xiaomi livestreamed its reinforcement-learning runs through a public dashboard over five days starting September 15, showing final costs of $854,044 for the smaller MiMo-V2.6-Flash model and $2,620,670 for Pro—roughly $3.5 million combined. Those figures cover only the RL stage; the company hasn't disclosed what pretraining cost. The published RL bills are rare in the industry: last year MiniMax said its 456-billion-parameter MiniMax-M1 incurred $534,700 in GPU rental for RL, while DeepSeek pegged its 671-billion-parameter R1 at $294,000. Xiaomi released the model weights under the permissive MIT license alongside a technical report, a 9-billion-parameter Qwen-based model for agentic RL research, and promised more than 7,000 task environments spanning software engineering, vulnerability reproduction, knowledge work and web development—though at the time of publication, only the three model releases appeared on Hugging Face.
Fuli Luo, who leads Xiaomi's MiMo team after working at DeepSeek, explained on X that her team spent nearly six months exploring how far RL could scale, increasing training volume, environment variety, and resources used to evaluate the model's output. Thomas Wolf, Hugging Face co-founder and chief science officer, called the effort an "Impressive level of openness on such a large run." According to Wolf, releasing high-quality open-source RL environments "is the most impactful thing anyone can do to push the open-source frontier right now," comparing it to sharing high-quality pretraining data but within the new reinforcement learning with verifiable rewards paradigm.
The report highlights how MiMo-V2.6 illustrates a distinction often blurred in AI development: open-weight versus open-source. Many "open" models amount to downloadable weights—the numerical values learned during training—while the production process remains closed. Meta has referred to its Llama models as open-source despite restrictions that prompted pushback from open-source advocates. MiMo-V2.6 goes further than most releases because its MIT license carries none of the conditions that models like Moonshot's Kimi K3 and Alibaba's Qwen3.8-Max impose on large commercial users, and if Xiaomi delivers the promised environments, outside researchers will have far more of the post-training process to examine and build upon. Environments become critical in RLVR because that approach depends on tasks whose outcomes can be checked automatically—whether code passes a test, for instance—making the training grounds themselves a core ingredient. For organizations evaluating whether to build on proprietary platforms or invest in open alternatives, the calculus now extends beyond model performance to the availability of reproducible training infrastructure and the licensing freedom to deploy without commercial restrictions.