Moonshot AI Unveils 2.8-Trillion-Parameter Kimi K3, First Open-Weight 3T Model; Musk Impressed

In July 2026, artificial intelligence company Moonshot AI released its next-generation model Kimi K3, with total parameters of 2.8 trillion, making it the world's first open-weight 3T-level model. Tesla CEO Elon Musk commented "Impressive" beneath a related benchmark report, his second public endorsement of the team. Four months earlier, Musk had praised on social media a technical paper titled "Attention Residuals" by the Kimi team, which sought to rewrite the core architecture that mainstream large language models have relied on for nearly a decade, and is one of the key supporting technologies for the K3 model. Among the paper's three equal-contribution co-first authors, Chen Guangyu, an intern from Shenzhen, was 17 years old and still in his final year of high school.

Born in 2009, Chen Guangyu attended an international school during his secondary education. In April 2024, he decided to pause the community service, coursework and other activities he had been accumulating for college applications, focusing instead on computer science and artificial intelligence. In March 2025, he began studying machine learning, downloading papers daily and using Gemini for question-based learning. In May of the same year, he shared a technical analysis of the DeltaNet linear attention mechanism on his personal social media account, receiving a reply from Yang Songlin, a PhD student at MIT, and was subsequently contacted by the CEO of US startup Tilde Research. In July 2025, he traveled alone to Palo Alto, California, for a seven-week internship at Tilde Research, participating in the completion of the MoMoE memory-optimized mixture-of-experts model project.

After the internship, Chen received an invitation from Moonshot AI and joined the Kimi expansion team. From January to March 2026, he completed the AttnRes research together with algorithm researcher Su Jianlin and Soochow University doctoral student Zhang Yutong. Su documented the division of work in his personal blog: he proposed the "seed idea" and conducted small-scale experiments, while Zhang and Chen subsequently proposed the Block Attention Residuals design, which divides layers into multiple blocks and compresses block information, reducing computation and communication costs while preserving effectiveness. After joining Moonshot AI, one of Chen's earliest tasks was accelerating the KDA (Kimi Delta Attention) computation kernel, giving him contributions to the development of both core architectures of Kimi K3.

Chen is not a "prodigy" in the traditional sense, holding no Olympiad medals or grade-skipping experience. In February 2025, at a middle school student programming hackathon in Beijing, he showcased his "human third mechanical assistive hand" concept and met judge Dong Kehan, later being selected for the latter's future leaders program for teenagers aged 15 to 17. Through Dong's introduction, Chen met DeepSeek researcher Yuan Jingyang, and under his guidance gradually built up his understanding of machine learning by studying papers and tracking open-source projects. In July 2025, he shared his learning methods on social media, saying he learned by understanding paper algorithms and reading Triton code on FLA, searching for relevant knowledge when encountering unfamiliar syntax. He once recorded in a blog that, in early 2025, to combat insomnia, he tried recording the relationship between his roommate's snoring frequency and factors such as exercise and diet in a table, training a predictive model that achieved 70% accuracy.

Chen's personal social media shows that he opened a video account in March 2025, producing under an hour of content daily and gaining 170,000 subscribers in four days. Moonshot AI officially released Kimi K3 in July 2026, with Chen participating in core architecture research and development as an intern, and his experience has been publicly reported.