Kimi K3
Also known as: Kimi K3 open weights, Moonshot K3, K3 model
Kimi K3 landed on July 16, 2026 via API and consumer apps, with full open weights published July 26. At 2.8 trillion parameters it is the largest open-weight model ever released and the first to reach what researchers are calling the 3T-parameter class. That matters because it puts a model competitive with closed frontier labs in the hands of anyone who can run the hardware or pay a hosting provider.
Under the hood it uses Moonshot's Kimi Delta Attention (KDA), a hybrid linear attention mechanism designed to handle very long sequences more efficiently, plus a Stable LatentMoE framework that activates 16 of 896 experts per inference step. The result is roughly 2.5x the scaling efficiency of its predecessor, Kimi K2. Context window is 1 million tokens, and it ships with native multimodal understanding.
For builders, K3's practical significance is threefold. First, it competes with closed-model benchmark scores on tasks like web research and coding. Second, as an open-weight model it can be self-hosted, fine-tuned, or routed through providers like Together AI and Modal with no vendor lock-in. Third, its release accelerated a broader conversation about what frontier open weights mean for deployment risk, data sovereignty, and the speed at which Chinese labs are closing the capability gap with American ones.