DEV Community

PhoenixWang
PhoenixWang

Posted on Originally published at en.wikipedia.org

The Long-Context Thesis: Yang Zhilin and the Academic Road to Kimi K3

When Kimi K3 landed in July 2026 — 2.8 trillion parameters, open weights, a 1-million-token context window, GPQA Diamond 93.5 — the PHLX Semiconductor Index slid into a bear market.

The two posts that best explain the model came from a university professor.

"He is absolutely brilliant," Ruslan Salakhutdinov wrote on X of his former PhD student Yang Zhilin, 34, founder of Moonshot AI. In another, he recalled Yang's words: if he didn't at least try starting his own company, he would regret it for the rest of his life.

Yang Zhilin at a Beijing conference

Three Tickets to Tsinghua

Born in Shantou in 1992, Yang came to coding late: a year of training won the provincial informatics olympiad — alone enough for Tsinghua admission. He passed Tsinghua's own exam too, then sat the gaokao anyway — 667, Shantou's top science score. Three tickets to one university.

Thermal engineering first. A Murakami novel flipped the path — one character was a gifted programmer — and Yang switched to computer science: "the most important and correct decision I've ever made." His band Splay, named after the splay tree, included Zhou Xinyu and Wu Yuxin, his future co-founders at Moonshot.

The Long-Context Thesis

At CMU he co-authored XLNet and Transformer-XL, finished the PhD in four years, and returned to China over offers from Apple, Stanford and MIT.

"If you can solve it with scale, don't solve it with a new algorithm. The new algorithm's value is to enable better scaling," Yang has said. His scaling target: context length.

Kimi K2 architecture and benchmarks, official report

In October 2023, Kimi launched with a 200,000-Chinese-character lossless window, a consumer-AI record; five months later, 2 million characters — a 10x jump. Yang's logic, March 2024: "From word2vec to RNN, LSTM, to Transformer, every architectural evolution has essentially been an improvement in effective, lossless context length."

Papers, Not Just Products

Kimi K2 (July 2025): 1.04T parameters, 32B active, 384 experts, 15.5T pretraining tokens, zero loss spikes. Top open model on LMSYS Arena.

Kimi K3 went further: the first open 3T-class model. Vercel's Guillermo Rauch: "the first time that an open model is ahead of all proprietary ones" on a web-engineering benchmark.

Kimi K3 benchmarks, official model card

The Cost of Speed

Capital tracked the thesis: worth just over $4 billion in December 2025, Moonshot hit $20 billion in a Meituan-led round by May, then $35 billion by July, with pre-IPO talks targeting $50 billion. A sevenfold re-rating in seven months.

Speed has a price. In November 2024, five investors from Yang's first startup Recurrent AI filed arbitration in Hong Kong over waivers they say he never obtained. GSR's Allen Zhu cited "uncleared liability"; Yang replied he had "gone through all necessary formalities." Still no settlement. After K3, OSTP director Michael Kratsios accused Moonshot of obtaining Nvidia chips despite export restrictions and distilling Anthropic's models.

The Drummer's Moon

Yang frames AGI as "not a staircase where you climb to a step and suddenly arrive. It is a direction." Asked about failure, he says he accepts it: "This has already completely changed my life, and I'm very grateful."

The Shantou kid who once wrote a song about getting rich overnight now runs a company named after a Pink Floyd album — and the open 2.8T model that made Silicon Valley recount its lead. His mentor's verdict holds: he was right.

Top comments (0)