Xiaomi open-sources top-scoring AI models amid Anthropic plagiarism accusations

Xiaomi released MiMo-V2.6 this week, a suite of open-weight AI models that has claimed the top spot among open-weight models on third-party benchmarks — arriving just days after Anthropic publicly accused the company of using Claude to extract training data.

The release, which Xiaomi made available on September 21-22, includes MiMo-V2.6-Pro, MiMo-V2.6-Flash, and a smaller 9-billion-parameter research model called MiMo-V2.6-Distill-Qwen-9B, all published under MIT licenses on Hugging Face. Pro features 1.02 trillion total parameters with roughly 42 billion active per query, while Flash offers 310 billion parameters with 15 billion active, both supporting text, image, audio, and video inputs with a one-million-token context window.eweek+2

Top Open-Weight Scores at a Fraction of the Cost

MiMo-V2.6-Pro scored 46 on the Artificial Analysis Intelligence Index, making it the highest-rated open-weight model tested by the independent benchmarking firm. According to VentureBeat, Flash is available via API at $0.14 per million input tokens and $0.28 per million output tokens, while Pro costs $0.435 and $0.87 respectively — a fraction of the price of leading proprietary models like Anthropic's Claude Opus 5, which charges $5 and $25 per million tokens.basic-tutorials+2

Beyond the model weights, Xiaomi published more than 7,000 reinforcement-learning task environments and a full training framework. The company reported that its RL training cost approximately $2.62 million for Pro and $850,000 for Flash.eweek

Anthropic's Distillation Accusations Loom Over the Launch

The release lands in a charged atmosphere. On September 10, Anthropic published a threat intelligence report accusing seven China-based AI labs — including Xiaomi — of conducting large-scale "distillation attacks" against Claude. Anthropic alleged that Xiaomi, under internal reference GTG-16008, forwarded more than 400,000 user conversations to Claude via intermediary tools between March and April 2026 to extract training data. Across all seven named companies, which also include Alibaba ↘ 4.55%, DeepSeek, and Moonshot AI, Anthropic documented nearly 200 million exchanges it classified as unauthorized distillation.qz+2

The Wall Street Journal reported that Chinese AI companies used networks of intermediary platforms to relay their users' queries to Claude, with some transmissions including sensitive data such as location information and passwords. Xiaomi has not publicly responded in detail to the allegations.wsj

What It Means for the Open-Weight Race

The timing has drawn scrutiny, as Anthropic's report preceded the MiMo-V2.6 launch by roughly two weeks. For enterprise adopters, the models themselves are publicly downloadable and independently testable. However, as eWeek noted, several benchmarks Xiaomi cites — including SWE-bench Verified and DeepSWE v1.1 — have been flagged by Epoch AI for scoring and contamination concerns, limiting the conclusions that can be drawn from those results alone.basic-tutorials+1

Whether the distillation allegations affect adoption will depend on how organizations weigh Xiaomi's competitive pricing and open licensing against unresolved questions about how the models' training data was sourced.