{"slug": "xiaomi-s-mimo-v2-6-tops-open-ai-charts-amid-anthropic-distillation-claims", "title": "Xiaomi's MiMo-V2.6 Tops Open AI Charts Amid Anthropic Distillation Claims", "summary": "Xiaomi published MiMo-V2.6 on Hugging Face on September 22, an open-weights mixture-of-experts model that scored 46 on Artificial Analysis's Intelligence Index, ranking first among the 114 open-weight models the firm tracks and level with Grok 4.7, just seven points behind Claude Fable 5.1 and GPT-6 Astra. Xiaomi says the reinforcement learning run behind MiMo-V2.6-Pro took under six days and cost about $2.62 million, with the smaller Flash model reportedly costing around $850,000, and released the weights, training framework, a 9B distilled version, and more than 7,000 RL environments under an MIT license. The release came less than two weeks after Anthropic's September 10 threat intelligence report, Case GTG-16008, alleged Xiaomi routed more than 400,000 conversations through Claude over 20 days in March and April 2026 using over 1,500 accounts, naming seven China-based labs with a combined tally of roughly 190 million exchanges.", "body_md": "*Xiaomi's new MiMo-V2.6 just became the highest-scoring open-weights AI model in the world, built for roughly $2.62 million. It shipped less than two weeks after Anthropic accused Xiaomi of running a Claude distillation campaign to help build it.*\n\nXiaomi published MiMo-V2.6 on Hugging Face on September 22. The numbers are hard to ignore. The flagship Pro model scored 46 on Artificial Analysis's Intelligence Index, the benchmark most of the industry now treats as the scoreboard, putting it first among the 114 open-weight models the firm tracks, level with Grok 4.7. According to Artificial Analysis, it trails Claude Fable 5.1 and GPT-6 Astra by just seven points - a gap that used to separate open models from the frontier by a much wider margin. Architecturally, it's a mixture-of-experts model: 1.02 trillion total parameters, only 42 billion active at once, wrapped in a 1 million-token context window that handles text, image, video, and audio.\n\nHere's the number that actually got people talking. Xiaomi says the reinforcement learning run behind MiMo-V2.6-Pro took under six days and cost about $2.62 million, according to a breakdown by Mixed News. Flash, the smaller sibling, reportedly cost around $850,000. Frontier labs spend hundreds of millions on runs like this. Xiaomi didn't just publish a model. It published a technical report, the full training framework, a smaller 9B distilled version, and more than 7,000 reinforcement learning environments with automatic graders, all under an MIT license that lets anyone download, fine-tune, and run the weights for free. On Xiaomi's own API, Pro runs $0.435 per million input tokens and $0.87 per million output tokens, with Flash priced at $0.14 and $0.28, according to pricing listed on OpenRouter.\n\n## The distillation allegations\n\nAnthropic has an answer, and it isn't flattering. On September 10, Anthropic's threat intelligence team published a report on what it calls Case GTG-16008: an industrial-scale campaign in which Xiaomi allegedly routed more than 400,000 conversations through Claude over 20 days in March and April 2026. Engineers used more than 1,500 distinct accounts and proxy services to mask where the traffic came from, Anthropic says, then replayed real user conversations and coding sessions from Xiaomi's own MiMo chatbot through Claude using coding harnesses called OpenClaw and OpenCode. The users whose sessions got relayed never agreed to any of it.\n\nXiaomi wasn't alone. Anthropic's report names seven China-based labs, Alibaba, Moonshot AI, DeepSeek, Zhipu, Xiaomi, MiniMax, and SenseTime, and puts their combined tally at roughly 190 million exchanges. Alibaba accounts for most of that on its own, more than 151 million interactions. DeepSeek ran a tighter version of the same play, cramming 12 million distillation attempts into 14 days in July, Anthropic says. Moonshot AI, the maker of the Kimi models, and Zhipu also appear on the list, though Anthropic's report gives fewer specifics on their individual campaigns.\n\n[This Week in AI: Robots Take a Punch, Assistants Get Long-Term Memory, and Models Keep Getting Cheaper](https://startupfortune.com/this-week-in-ai-robots-take-a-punch-assistants-get-long-term-memory/)\n\nA dense week across robotics, model releases and AI research: humanoid robots duke it out, ChatGPT gets long-term memory, and frontier-adjacent AI keeps getting cheaper and faster. - [how to build robots that handle physical impacts](https://startupfortune.com/this-week-in-ai-robots-take-a-punch-assistants-get-long-term-memory/) - [AI models getting cheaper and more accessible features](https://startupfortune.com/this-week-in-ai-robots-take-a-punch-assistants-get-long-term-memory/)\n\nThis isn't even the first warning. Anthropic and OpenAI jointly flagged similar extraction campaigns by Chinese labs back in February, according to CNBC. What changed by September is the scale, and the fact that a model partly built on one of these campaigns just topped the leaderboard.\n\n## What the silence means\n\nXiaomi hasn't responded publicly to Anthropic's allegations.\n\nThat silence says something on its own.\n\nFrankly, the open weights don't settle the ethics question, they just move it. Publishing the training code and the RL environments lets any researcher check whether MiMo-V2.6 stands on its own, and that transparency is real and worth crediting. It doesn't answer whether the data that trained it was gathered with consent. You can verify a model's architecture from its published weights. You can't verify what happened inside 400,000 conversations that, according to Anthropic, were never meant to leave the company that hosted them.\n\nAnthropic's report doesn't call for MiMo-V2.6 to be pulled or blocked. It logs the case, names the lab, and moves on to the next one, which for now is exactly how this fight over training data is getting resolved: quietly, in a compliance document, while the open-weights leaderboard keeps moving underneath it.\n\n**Also read:** [GoCardless Used an AI Agent to Move Real Money From a UK Bank Account](https://startupfortune.com/gocardless-used-an-ai-agent-to-move-real-money-from-a-uk-bank-account/) • [Why Does My AI Coding Agent Bill Keep Increasing Overnight](https://startupfortune.com/why-does-my-ai-coding-agent-bill-keep-increasing-overnight/) • [America's Trillion-Dollar Chip Boom Is Short 157,000 Workers It Doesn't Have](https://startupfortune.com/americas-trillion-dollar-chip-boom-is-short-157000-workers-it-doesnt-have/)\n\n[Bernie Sanders Proposes 20-Year Prison Terms to Ban Superintelligent AI](https://startupfortune.com/bernie-sanders-proposes-20-year-prison-terms-to-ban-superintelligent-ai/)\n\nSen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act on September 3, proposing a permanent ban on superintelligent AI with 20-year prison terms and a corporate death penalty for violators. The bill lands just as Trump unveils an \"AI Force\" and calls superintelligence fears a hoax, setting up a sharp clash in... - [how to regulate superintelligent AI development legally](https://startupfortune.com/bernie-sanders-proposes-20-year-prison-terms-to-ban-superintelligent-ai/) - [Bernie Sanders ban superintelligent AI 20 year prison](https://startupfortune.com/bernie-sanders-proposes-20-year-prison-terms-to-ban-superintelligent-ai/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.", "url": "https://wpnews.pro/news/xiaomi-s-mimo-v2-6-tops-open-ai-charts-amid-anthropic-distillation-claims", "canonical_source": "https://startupfortune.com/xiaomis-mimo-v26-tops-open-ai-charts-amid-anthropic-distillation-claims/", "published_at": "2026-09-23 17:36:11+00:00", "updated_at": "2026-09-23 17:58:43.264649+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-safety", "ai-policy", "ai-research"], "entities": ["Xiaomi", "MiMo-V2.6", "Anthropic", "Claude", "Artificial Analysis", "Alibaba", "DeepSeek", "Moonshot AI"], "alternates": {"html": "https://wpnews.pro/news/xiaomi-s-mimo-v2-6-tops-open-ai-charts-amid-anthropic-distillation-claims", "markdown": "https://wpnews.pro/news/xiaomi-s-mimo-v2-6-tops-open-ai-charts-amid-anthropic-distillation-claims.md", "text": "https://wpnews.pro/news/xiaomi-s-mimo-v2-6-tops-open-ai-charts-amid-anthropic-distillation-claims.txt", "jsonld": "https://wpnews.pro/news/xiaomi-s-mimo-v2-6-tops-open-ai-charts-amid-anthropic-distillation-claims.jsonld"}}