Plagiarism Allegations Between Stanford's Llama3‑V and China's MiniCPM‑Llama3‑V 2.5 Model

After a wave of accusations that Stanford's Llama3‑V multimodal model copied the architecture and code of the Chinese MiniCPM‑Llama3‑V 2.5 model, the Llama3‑V team publicly admitted the plagiarism, while two undergraduate members and a key contributor from the University of Southern California, Mustafa Aljadery, disappeared.

"We hope to get a statement from Aljadery, but he has been unreachable since yesterday," the post reads.

The apology tweet, issued by Siddharth Sharma and Aksh Garg, emphasizes regret but offers little substantive remediation, prompting criticism from Stanford AI Lab director Christopher Manning, who accused the team of refusing to acknowledge their mistake.

New evidence shows that the weight differences between Llama3‑V and MiniCPM‑Llama3‑V 2.5 follow a near‑zero‑mean Gaussian distribution with a standard deviation of 1.4e‑3, suggesting that Llama3‑V may have simply added low‑variance noise to the original weights.

Additional allegations claim that Aljadery previously authored a plagiarized computer‑network‑design textbook, which has since been taken down.

MiniCPM‑Llama3‑V 2.5, built on SigLip‑400M and Llama3‑8B‑Instruct, achieves an average OpenCompass score of 65.1, surpassing proprietary models such as GPT‑4V‑1106, Gemini Pro, Claude 3, and Qwen‑VL‑Max, and demonstrates strong OCR performance (700+ on OCRBench) and a low hallucination rate (10.3% on Object HalBench).

"We are confident Llama3‑V is a shell of our MiniCPM‑Llama3‑V 2.5," said Liu Zhiyuan, chief scientist of MiniCPM.

Commentators, including DeepMind researcher Lucas Beyer and Hugging Face community lead Omar Sanseviero, argue that Chinese open‑source models are often overlooked despite achieving competitive results, citing examples such as Qwen, Yi, DeepSeek, and others.

"The community has been ignoring the Chinese machine‑learning ecosystem, even though they produce impressive large‑language, vision, and diffusion models," Sanseviero noted.

Overall, the controversy highlights both ethical concerns about model plagiarism and the broader issue of recognition for Chinese contributions in the rapidly evolving AI landscape.