Huawei cloned Qwen and DeepSeek models, claimed as own

10 months ago

Huawei's Pangu model was accused of plagiarism from Qwen 2.5, leading to internal controversy.
A Huawei researcher revealed that another team within the company rebranded Qwen and DeepSeek models as their own, claiming they were trained from scratch.
The Pangu team faced numerous challenges, including inefficient tokenizers, failed model training, and intense pressure to deliver results.
Despite hardships, the team successfully developed the Pangu V3 model from scratch, achieving stable performance on Ascend hardware.
Internal politics and mismanagement led to talent attrition, with many engineers leaving for companies like ByteDance, DeepSeek, and Tencent.
The whistleblower expressed shame and frustration over the internal fraud and resigned, calling for transparency and ethical practices in AI development.

Hasty Briefsbeta