Kimi's 17-year-old Attention Residuals co-author leaves Moonshot AI Other

Chen Guangyu, a 17-year-old researcher at Moonshot AI, said on October 15 that he has left the company behind the Kimi models. He is listed first among three equal-contribution authors of "Attention Residuals", a paper whose method Moonshot built into its 2.8-trillion-parameter Kimi K3 model. He says he is heading to the San Francisco Bay Area. He has not named a new employer.

Chen joined the Kimi team as an intern in November 2025, at 16, while still in high school. Chinese media report that Moonshot hired him as an intern because he was under 18. He finished high school in June 2026.

What he said

• In a post on X, reported by SBS News, he wrote that he had "left Moonshot" and that "the past year has been the best of my life".
• He thanked the company for hiring him when he was 16.
• Anthropic researcher Sholto Douglas replied with a job pitch. Chen answered that he is "heading to the Bay Area".
• Asked whether he would join Anthropic, he replied "That is not happening", SBS reports.
• He did not give a reason for leaving or name a next job.

The paper

• Title: "Attention Residuals", by the Kimi Team, posted on arXiv on 16 March 2026.
• Chen is listed first. The paper marks him, Yu Zhang and Jianlin Su as equal contributors.
• The idea: standard residual connections add up every layer's output with fixed weights. Attention Residuals lets each layer use attention to choose which earlier layers to draw from.
• The authors tested it on a 48B-parameter Kimi Linear model trained on 1.4T tokens.
• Chinese outlets, including Huxiu, report that Elon Musk reposted and praised the paper.

Where it was used

• Moonshot's model card describes Kimi K3 as "a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes)".
• The card lists 104B activated parameters. K3 was released in July 2026.

What is not confirmed

• A widely shared claim that he earns 70 million yuan a year has no source. Chinese headlines used a much smaller figure, taken from an investor's post about an unnamed 17-year-old with job offers. Chen has not confirmed any offer or pay.
• A claim that he joined Anthropic's Claude team contradicts his own public denial.

Sources

• Attention Residuals (arXiv:2603.15031)
• Kimi K3 model card, Moonshot AI on Hugging Face
• SBS News: China's teenage AI prodigy leaves Moonshot AI
• Seoul Economic Daily: Teen prodigy behind Kimi breakthrough leaves Moonshot AI
• Sina (Southern Metropolis Daily)
• Huxiu