Xiaomi has officially launched its MiMo-V2.6 model suite, led by MiMo-V2.6-Pro, a 1.02-trillion parameter mixture-of-experts model that ranks at the top of open-weights benchmarks. Scoring 46 on the Artificial Analysis Intelligence Index, Pro costs $0.435 per million input tokens and $0.87 per million output tokens. Xiaomi attributes the model’s capabilities to an expanded reinforcement learning phase costing roughly $2.62 million over six days.

In tandem with the release, Xiaomi opened up its complete RL framework, including technical reports, training code, auto-graders, and thousands of synthetic and real-world task environments. However, the launch arrives shortly after Anthropic accused Xiaomi and six other Chinese labs of illegal distillation. Anthropic reported tracking over 400,000 exchanges where Xiaomi allegedly siphoned outputs from Claude to train its internal models.

The release illustrates the intense price-performance competition coming from Chinese AI labs using aggressive RL scale-ups. However, growing friction around cross-border distillation and intellectual property access could lead to stricter API rate limits and increased security measures by Western frontier labs.

Why it matters

  • Cost-efficiency gains: Xiaomi’s Pareto-efficient pricing provides extremely cheap frontier-class capabilities for enterprise agentic workloads.

  • Open-source tooling: Access to Xiaomi’s RL training framework and task environment open-sources advanced post-training capabilities for researchers.

  • IP and compliance risks: Enterprise users must track supply chain provenance as frontier labs increase scrutiny on model scraping and distillation.

Source: the-decoder.com