Xiaomi's latest AI model MiMo-V2.6-Pro has claimed the top position among openly available large language models while pricing aggressively below competitors. The achievement came through intensive reinforcement learning that consumed $2.62 million in compute costs. However, the accomplishment carries significant controversy: Anthropic publicly accused Xiaomi of extracting training data directly from Claude to build the model.
The MiMo-V2.6-Pro represents a shift in the open-source AI landscape. Xiaomi positioned the model as an affordable flagship alternative to proprietary systems and competing open models. By reaching top-tier performance benchmarks while maintaining lower costs than enterprise-grade systems, the Chinese smartphone manufacturer signaled its intent to compete in the foundation model space beyond its hardware ecosystem.
The $2.62 million reinforcement learning investment reflects the resource requirements for pushing open models toward performance parity with industry leaders. Reinforcement learning from human feedback (RLHF) and related techniques require substantial computational infrastructure to train models on reward signals rather than just supervised data. Xiaomi's willingness to spend heavily on this phase suggests the company views AI capabilities as central to its competitive strategy going forward.
Anthropic's allegation presents a different story. The company claims Xiaomi systematically used Claude's outputs as training material for MiMo-V2.6-Pro. This practice, sometimes called "distillation" when legitimate, becomes problematic without explicit permission or proper attribution. Anthropic built Claude through its own expensive training processes and research investments. Using Claude's outputs to train competing models without consent undermines the economics and incentives for companies investing in frontier AI development.
The accusation touches a raw nerve in open-source AI circles. The community celebrates transparency and accessibility, but Anthropic's complaint highlights a tension: when models become good enough, their outputs become valuable training material. Companies can theoretically extract knowledge from public-facing APIs without direct agreement, creating a form of parasitic competition. Xiaomi gains the benefit of Claude's training investments without bearing those costs.
Xiaomi has not publicly responded to the allegations as of publication. The company's silence leaves several questions unresolved. Did Xiaomi knowingly use Claude outputs at scale, or did it occur as part of a broader data collection process? Did the company consider this practice acceptable within open-source norms? Will Xiaomi modify its approach or defend the decision?
The pricing undercut matters for market dynamics. If Xiaomi can achieve top-tier performance at lower cost through data extraction from competitors, it creates perverse incentives. Other companies might follow the same playbook, potentially degrading the willingness of frontier labs to make models accessible through APIs or free tiers. The outcome could paradoxically push AI development toward closed systems as companies protect against such usage.
MiMo-V2.6-Pro's performance gains remain real regardless of the training controversy. The model delivers measurable improvements on standard benchmarks. Whether those improvements justify Xiaomi's methods from an ethical or legal perspective remains contested. The dispute exemplifies broader questions about fair competition in AI: what constitutes acceptable training data sourcing when models become public products, and who bears responsibility when those boundaries blur.