@turingbook: Let the insiders tell the history. The earliest large model research in China was basically related to the Beijing Academy of Artificial Intelligence (BAAI). The first officially released large model was Qingyuan CPM (November 2020, 2.6 billion parameters) supported by BAAI and the team of Liu Zhiyuan from Tsinghua University. https://zhuanlan.zhihu.com/p/2981380...
Summary
The article reviews the early development history of large models in China, pointing out that BAAI supported the earliest Qingyuan CPM (2020) and Wudao 1.0 (2021), and corrects the claim that Huawei Pangu was the first domestic large model.
View Cached Full Text
Cached at: 06/17/26, 05:49 AM
Let the insiders tell the history. The earliest large model research in China is largely tied to the Beijing Academy of Artificial Intelligence (BAAI). The first officially released large model was Qingyuan CPM (November 2020, 2.6 billion parameters) from BAAI supporting Zhiyuan Liu’s team at Tsinghua University https://zhuanlan.zhihu.com/p/298138008 , which later evolved into OpenBMB and remains one of the best on-device models today. Next came the Wudao 1.0 model series (March 2021), led by BAAI with Professor Tang Jie as the principal investigator. This series included the earliest version of GLM (GLM paper corresponding authors were Tang Jie and Yang Zhilin). Other early works include Baidu’s ERNIE 3.0 (July 2021); Alibaba’s M6 (March 2021), which was a collaboration with Tang Jie’s team (Lin Junyang was first author, Tang and Yang Hongxia [then at Alibaba] were corresponding authors), later evolving into Qwen. Huawei’s Pangu had two versions: one from Huawei Cloud (Tian Qi) in collaboration with Yang Zhilin’s team, and another from the Noah’s Ark Lab, but neither was sustained. There may be omissions; feel free to add.
Xudong Han (@Xudong07452910): I happened to scroll past Richard Yu saying at a Huawei press conference that Pangu was the first large model in China, so I checked the timeline.
The PanGu-α paper was posted on April 26, 2021, and already called itself a Large-scale Pretrained Language Model in the title and abstract, with up to 200B parameters. Baidu’s ERNIE 3.0 came on July 5, 2021, so in terms of timeline it really was a step behind.
Similar Articles
@seclink: Recently, there was a Huawei press conference and the Pangu large model is making a comeback. Open-source weights on June 20. I guess Claude and GPT will have to step aside. Is the world's best large model about to be born?
Huawei press conference announced that the Pangu large model will make a comeback, planning to open-source weights on June 20, claiming it may surpass Claude and GPT to become the world's best.
@Fenng: The WeLM large model developed by the WeChat team has likely entered the top tier of domestically produced large models.
The WeLM large model developed by the WeChat team is considered to have entered the top tier of domestically produced large models.
@MaxForAI: After reading the remarks of former Qwen researcher Hongyi Yuan @yiguyuan20, I finally understand why the foundational models from major Chinese tech companies can't make real progress... So pointing out mistakes or telling the truth requires seniority?
Remarks by former Qwen researcher Hongyi Yuan reveal a culture of seniority in foundational model R&D at major Chinese tech companies, which stifles newcomers from raising objections or innovating, hindering model breakthroughs.
@shenxiankk: Stop expecting others to spoon-feed you a ready-made meal — this guide walks you hand-in-hand from scratch to fully jailbreak a large model yourself. Pliny, the jailbreak expert named one of TIME's 100 Most Influential People in AI, has open-sourced his core toolkit OBLITERATUS: it's not a finished model but an entire industrial-grade…
Pliny (named to TIME's 100 Most Influential People in AI) has open-sourced the toolkit OBLITERATUS — an industrial-grade pipeline that eliminates a large model's refusal mechanism without retraining by locating and surgically removing the internal refusal direction (abliteration), while preserving its language capabilities. The toolkit offers multiple extraction strategies including PCA, mean difference, sparse autoencoder decomposition, and whitened SVD, along with a Gradio interface and Python API, and ships with 15 deep-analysis modules. Each run also anonymously contributes benchmark data for crowdsourced research.
@leo114119: It’s confirmed: from now on, all Chinese open-source models must have their cybersecurity capabilities stripped out before they can be publicly released: (K3.1 Pro is probably the first.... The upcoming GLM 5.5 /Qwen4 will probably have theirs stripped out when they’re open-sourced too..) Fortunately, previously open-sourced models like GLM 5.3 and KIMI K3 should all…
The user claims that Chinese open-source models will have to strip out cybersecurity capabilities before they can be released in the future, speculates that future models such as K3.1 Pro, GLM 5.5, and Qwen4 will all be affected, and recommends saving existing open-source models with their full capabilities intact.