Xiaomi has released MiMo-V2.6-Pro, a 1.02-trillion-parameter mixture-of-experts model that activates 42 billion parameters per request. Artificial Analysis gives it 46 points on its Intelligence Index, currently placing it ahead of other openly available models in that ranking.
API prices are listed at $0.435 per million input tokens and $0.87 per million output tokens. Artificial Analysis estimates about $0.13 per benchmark task. Xiaomi attributes the performance jump to reinforcement learning across coding, cybersecurity, office and web-design environments. It says the main training run lasted under six days and cost $2.62 million; a smaller Flash model cost $0.85 million.
The company is releasing its framework, technical report, a smaller training model and thousands of automatically graded tasks. An UltraSpeed variant promises up to 20 times faster output, though that figure is a vendor claim and needs workload-specific testing.
The release also carries a dispute. Anthropic alleges Xiaomi was among several Chinese labs that improperly used Claude outputs for training, identifying more than 400,000 exchanges associated with Xiaomi in one case. Xiaomi’s open tooling does not resolve that allegation, and benchmark leadership does not establish how the model performs across untested production tasks.