AINEWS
Tang Jie signed the form nominating Yang Zhilin for a Tsinghua prize. Z.AI and Moonshot AI now compete with Anthropic and OpenAI on public benchmarks, and Anthropic says they used its models to get there.
Tang Jie and Yang Zhilin were teacher and pupil more than a decade ago, with dreams of building a machine that could think like a human. Yang went on to lead Moonshot AI. Tang co-founded Z.AI. The teacher-student pair now sit at the center of a Chinese open-source AI stack that the Hindustan Times profile reports as competitive with US labs in published benchmarks, and that Anthropic has accused of being partly built by distilling its own models.
The institutional mechanism is older than the "China caught up" story usually allows. A 2022 ACL paper on the GLM pretraining framework, co-authored by Tang, Yang, and others, and the GLM-130B preprint put bilingual open-weight pretraining on the research map years before the current model release cycle. The Hindustan Times profile traces Tang's research career back roughly a quarter-century and describes the Tsinghua computer science department as a hub that has spun out a generation of Chinese AI researchers, many of them trained in the US before returning. Z.AI and Moonshot AI are each reported in that profile to be valued at tens of billions of dollars, though those figures are narrative estimates rather than primary funding records.
The model releases themselves follow the same open-source pattern. The most recent test of that stack is Kimi K2, a 1-trillion-parameter mixture-of-experts model from Moonshot with 32 billion parameters active per token, pre-trained on 15.5 trillion tokens. Mixture-of-experts is an architecture in which only a fraction of the model's parameters fire on any given input, which lets labs train larger systems without paying full inference cost on every question. Moonshot's technical report records 66.1 on Tau2-Bench, 76.5 on ACEBench (English), 65.8 on SWE-Bench Verified, and 47.3 on SWE-Bench Multilingual, scores the company itself published, and that the open-weight release on Hugging Face makes reproducible for outside researchers. Z.AI's earlier open-weight models, also available publicly, sit in the same lineage.
That openness is what made the dispute visible. Distillation, in machine-learning terms, is the practice of training a new model by asking an existing one many questions and learning from its answers, a shortcut that some US labs argue violates their terms of service. Anthropic has publicly accused both Z.AI and Moonshot of running large-scale distillation against its models; the allegation is a company-side claim, not a judicial or independent finding, and it sits alongside the open-source lineage rather than replacing it. Researchers can verify Moonshot's benchmark claims because the weights are out; that same openness also leaves a paper trail that makes Anthropic's accusation possible to investigate at all.
US export controls and IP enforcement both presume a primarily adversarial relationship. The Tsinghua record points to a different mechanism: long-running public research, open bilingual pretraining, and the steady return of US-trained Chinese researchers to institutions that then spin out well-funded startups. The distillation accusations address one channel; the open bilingual pretraining channel runs underneath it. Whether Anthropic's allegations result in any independent finding is now the open legal question. The technical question is simpler: the next Kimi K2 release is already in the pipeline, and the open weights are already on Hugging Face.