Moonshot AI just dropped Kimi K3, a new AI model that can process up to 1 million context tokens in a single prompt. That’s enough to digest an entire large codebase or a full-length book without breaking a sweat. The model is now live across web, apps, and APIs.
What Kimi K3 actually brings to the table
The headline number is the context window: 1 million tokens. To put that in perspective, the previous K2 series models supported context windows of roughly 256K to 262K tokens. K3 represents roughly a four-fold expansion in how much information the model can hold in its working memory at once.
The model is expected to run on approximately 2 to 3 trillion parameters using a Mixture-of-Experts (MoE) architecture. MoE is the same approach that powers some of the most capable models in the industry, routing different queries to specialized sub-networks rather than firing every parameter for every task. The K2 series ran on about 1 trillion parameters, so K3 roughly doubles or triples that capacity.
Then there’s the Agent Swarm technology. K3 can coordinate up to 300 sub-agents working in parallel, enabling complex multi-step planning and execution.
