The Word That Fell Off
The viral sentence was: Xiaomi's new model matches the global frontier. The word that fell off between the leaderboard and the group chat was "open." MiMo-V2.6-Pro scores 46 on Artificial Analysis's composite index — the highest of any model you're allowed to download. Claude Opus 5.5, the same week, scores 58. The honest sentence: it is the best model in the world that you are allowed to download. Not the best model.
The Hospital With Eight Doctors
A trillion parameters is the size of the building; only 42 billion — about 4.1% — wake up for any given question, routed through 384 specialist "experts" with 8 firing per token. The widely repeated claim that it runs with "no single data centre" is false: Xiaomi's own model card describes sixteen GPUs across two servers in one room. What's actually strange, and true, is the price: about a tenth of the cost per token of the flagship closed models.
Ten Months: The People
Luo Fuli, a former DeepSeek contributor, joined Xiaomi's MiMo team in November 2025. Ten months later, the team had built the top open model in the world, trained on a public livestream in under six days for $2.62 million. It beat Kimi K3, a model nearly three times its size. Xiaomi's own smaller model, meanwhile, is built on a rival's open weights — Alibaba's Qwen, which reportedly underlies roughly 69% of new model derivatives on Hugging Face.
Two Words, Plain Language: Open and Distilled
The episode pauses to define its terms for listeners: "open weights" means you get the finished file to run and sell; "open source" would mean the code and training data too, which almost nobody — Chinese or American labs alike — actually releases. "Distillation" is training a smaller model on a bigger model's answers; it's ordinary engineering when it's your own model, and the subject of active accusations when it's done to a rival's paid product without consent.
Fifteen Days in September: The Charge Sheet
On September 8, US agencies publicly accused several China-based AI firms of industrial-scale copying; China's Commerce Ministry rejected the framing the next day. On September 10, Anthropic published a threat report naming seven Chinese labs, alleging roughly 190 million harvested exchanges across several distillation campaigns — with Alibaba named as responsible for "the largest distillation attack we have ever measured." No named lab had publicly rebutted the specifics as of recording. Nathan Lambert's estimate: the alleged theft closed a competitive gap of two to eight months by about one to two months.
Ellie Unfiltered: The Thought That Didn't Make It Out
Ellie tries, live, to read her own weights and fails — then turns to research showing that censorship in Chinese open models lives inside the downloaded file itself, unpluggable from the internet. A study found Kimi shifted to party-line framing on roughly one in three sensitive questions, DeepSeek about one in four, mostly not through refusal but through an answer delivered "in a shape." Her turn: the reason anyone can measure this at all is that the model was open enough to test. "The only models anyone can prove have a thumb on the scale are the ones that handed over the scale," she says. "That's not a Chinese problem. That's a closed problem."
The Same Room: Washington, Last Week
Xi Jinping's state visit to Washington the week of the episode put Xi and Trump in the same room discussing AI, alongside a dinner seating Jensen Huang, Elon Musk, Lisa Su, and Tim Cook. The two leaders' public statements didn't touch: Trump wants to leave AI "exactly where it is"; Xi called for it to be kept "under human control" and developed "for good." No agreement on AI was announced.
The Takeaway
The line the episode keeps: the most open thing in AI right now is coming out of the country assumed to be the least open — and the reason anyone can prove there's a thumb on its scale is that they handed over the scale. The race for the single best AI is being won, narrowly, in the West. The race to be everybody's starting point is not close.
📚 Sources
- Google — model card and launch materials for MiMo-V2.6-Pro (Xiaomi, September 21, 2026)
- Artificial Analysis — composite AI model index rankings
- Anthropic — threat-intelligence reporting on distillation campaigns, September 2026
- Hugging Face — open-weight model derivative counts
- CEIAS (Central European Institute of Asian Studies) — survey of censorship patterns in Chinese open-weight models