Moonshot AI launches 2.8T parameter Kimi K3 MoE model
Moonshot says its new Kimi K3 model brings a million-token context window, native multimodality and open weights planned for July 2026.
Entities: Zhilin Yang
Moonshot AI said on X that its new Kimi K3 model has 2.8 trillion parameters, a 1 million-token context window, native multimodality and new efficiency techniques aimed at long-horizon agentic coding. The company also said open weights are planned for July 2026. So far, the public details driving this story come from Moonshot's own announcement and posts reacting to it on X.
Combined views
290K
25 posts, first seen 6h ago
Commentary on X
On paper( I have not tested yet) Kimi K3 represents a serious architectural step forward by combining the efficient linear attention (Kimi Delta Attention) with smarter cross-layer residual routing (Delta Attention Residuals). The early leaderboard signals in coding/frontend/agentic arenas are impressive, especially for long-horizon work. It’s one of the most interesting open-weight frontier releases in a while not just bigger parameters, but meaningfully better mechanisms for the exact workloads-agentic coding, long context, self-improving workflows that matter most right now.
@rsalakhu @Kimi_Moonshot Such an inspiring story! Congratulations to the @Kimi_Moonshot team and Zhiling