Moonshot AI's Kimi K3, with 2.8 Trillion Parameters, Tops Coding Blind Test; Zhipu Shares Drop 42% in Two Days
Moonshot AI released Kimi K3, a fully open-source model with 2.8 trillion parameters. In an anonymous blind test voted by millions of users, K3 jumped from 18th to first place on a coding leaderboard. On the same day, Zhipu AI’s shares plunged 28%, followed by a further 19% decline the next session, for a two-day drop of about 42% and a loss of roughly HK$300 billion in market value. The model uses a KDA hybrid linear attention mechanism for 2.5x scaling efficiency. On OpenRouter, Chinese open-source model usage climbed from under 2% at end-2024 to over 50% in February and held above 40% in April.
On July 17, Moonshot AI released Kimi K3, an open-source model with 2.8 trillion parameters. In an anonymous blind test voted by millions of users, K3 rose from 18th in the previous generation to first place on a coding leaderboard. The same day in Hong Kong trading, Zhipu AI shares tumbled 28%, and fell a further 19% the next trading session, resulting in a two-day cumulative decline of about 42% and erasing approximately HK$300 billion in market capitalisation. The US semiconductor sector also suffered a sharp sell-off.
K3 adopts a KDA hybrid linear attention mechanism, achieving 2.5 times scaling efficiency at lower computational cost. This architectural innovation path, similar to DeepSeek’s MLA architecture, reflects how chip restrictions and open-source competition are driving engineering capabilities toward architectural advances.
On the global API aggregation platform OpenRouter, the share of calls to Chinese open-source models rose from below 2% at end-2024 to a peak of over 50% in February this year, and still held above 40% in April. Independent research institute METR’s measurements show that the time AI can perform a task independently has doubled every seven months since 2019, accelerating to just over four months this year. In an official K3 demonstration, the model worked continuously for 48 hours and completed the full design of a 4-square-millimetre chip containing 1.46 million standard cells, with timing closure achieved.