Xiaomi's open-source AI model achieves near-zero cost per token

Xiaomi's open-source AI model achieves near-zero cost per token

Xiaomi released MiMo-V2.6-Pro, an open-source AI model scoring 46 points on the Intelligence Index. It operates at $0.435 per million input tokens and $0.87 per million output tokens—prices that approach near-zero cost for mass adoption. The model contains 1.02 trillion parameters but activates only 42 billion per request, enabling efficient use. Xiaomi claims its performance improved through expanded reinforcement learning, achieving a 72.6/58.4 score boost on coding tests. It was trained on 7,000 real GitHub pull requests and synthetic environments, supporting software development, cybersecurity, and creative work.

The affordability stems from Xiaomi’s open model architecture and low-cost training—$2.62 million for the base version versus $0.85 million for a lighter variant. This pricing structure could let billions access AI tools for coding, design, and productivity without subscription fees. For the knowledge sector, it represents a potential shift toward free, high-capacity AI assistance where computational cost becomes negligible.

This matters because it directly lowers barriers to technical skills for underserved populations. However, Anthropic reports a single case study (GTG-16008) where 400,000 exchanges occurred between Xiaomi’s model and Claude over 20 days in March-April 2026. Xiaomi denies data siphoning, but Anthropic’s threat intelligence (Dec 2025–Aug 2026) highlights ongoing tensions about model data sharing. The model’s real-world impact depends on whether its open nature holds up under scrutiny and whether the low cost translates to actual usage beyond the training environment.

Source: The Decoder