Moonshot AI Unveils 2.8-Trillion-Parameter Open-Source Kimi K3; Zhipu Shares Plunge 42% in Two Days
Moonshot AI released Kimi K3, a fully open-source model with 2.8 trillion parameters, on July 17. In an anonymous blind-test coding ranking, K3 jumped from 18th to first place. Zhipu's shares fell 28% on the first trading day and 19% the next, erasing about HK$300 billion in market value. US semiconductor stocks also dropped sharply.
In the early hours of July 17, Moonshot AI released the Kimi K3 model, with 2.8 trillion parameters, fully open source. In a coding leaderboard based on anonymous blind-test voting by millions of users, K3 rose from 18th place in the previous generation to first place. On the same day, as Hong Kong markets opened, Zhipu's shares fell 28%, followed by a further 19% decline the next trading day, erasing approximately HK$300 billion in market value. US semiconductor stocks also suffered heavy losses.
K3 employs a KDA hybrid linear attention mechanism, achieving 2.5 times scaling efficiency at lower computational cost. This architectural innovation path is similar to DeepSeek's MLA architecture, reflecting a shift in engineering capability toward architectural innovation amid chip restrictions and open-source competition.
On OpenRouter, a global API aggregation platform, the share of calls to Chinese open-source models rose from less than 2% at the end of 2024 to a peak of over 50% in February this year, and remained above 40% in April. Independent research by METR shows that the duration for which AI can autonomously complete tasks has doubled every seven months since 2019, accelerating to a doubling every four months this year. In an official demonstration, K3 worked continuously for 48 hours to complete the full design of a 4-square-millimeter chip, containing 1.46 million standard cells, with timing closure achieved.