A Beijing-based AI startup just released the largest open-weight model the world has seen, and the ripple effects are already hitting tech stocks, chip makers, and the broader conversation around who actually controls the future of artificial intelligence.
Moonshot AI launched its Kimi K3 model on July 16, packing 2.8 trillion parameters into a natively multimodal architecture that handles text, images, and video. The model’s one-million-token context window puts it in rarefied company. Its API pricing, at $3 per million input tokens and $15 per million output tokens, is roughly half what comparable US models charge.
What Kimi K3 actually does
Kimi K3 is built on a hybrid architecture combining what Moonshot calls “Kimi Delta Attention” with a sparse Mixture-of-Experts setup. The model doesn’t fire all 2.8 trillion parameters for every query. It routes tasks to specialized sub-networks, which keeps compute costs manageable while maintaining performance across diverse workloads.
The model excels at long-horizon coding tasks, agentic workflows, and visual reasoning. Moonshot AI plans to release the full open weights on July 27, meaning developers worldwide will be able to fine-tune the model for their own applications.
