China’s Moonshot AI released Kimi K3 on Thursday, a 2.8-trillion-parameter open-weight model that benchmarks show trading blows with Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol at the top of the leaderboard.
The model, timed to land just ahead of the World AI Conference in Shanghai, is roughly 75% larger than DeepSeek’s V4 Pro and features a 1-million-token context window, native visual understanding, and an always-on reasoning mode called “thinking mode.” Full model weights are scheduled for open release on July 27.
Kimi K3 is built on two architectural innovations from Moonshot: Kimi Delta Attention, a hybrid linear attention mechanism, and Attention Residuals, which deliver consistent scaling gains. On the API side, it is compatible with the OpenAI SDK, priced at $3 per million input tokens and $15 per million output tokens.
The model scored third overall on the GDPval-AA v2 benchmark at 1,687, behind only Claude Fable 5 Max and GPT-5.6 Sol Max. It achieved state-of-the-art on BrowseComp with a score of 91.2 out of 100, and ranked first in four of eight real-world task automation benchmarks.
