Moonshot AI Releases Kimi K3, the World's Largest Open-Source Model

On July 17, 2026, Chinese AI lab Moonshot AI released Kimi K3, a model with 2.8 trillion total parameters. It is the largest open-source language model in the world. The company also announced that the full model weights will be released to the public on July 27, allowing any developer to study, modify, or run the system on their own hardware. K3 uses a sparse Mixture-of-Experts design, activating only 16 of its 896 expert networks for each token. A new attention mechanism, called Kimi Delta Attention, makes long-context inference up to 6.3 times faster. The model can also handle a one-million-token context window, large enough to read an entire novel in a single prompt. In independent tests, K3 topped the Arena AI leaderboard for front-end coding, beating the latest models from leading United States rivals. On the SWE Marathon benchmark, which measures long software engineering tasks, K3 scored 42.0, higher than every Western competitor. Developers say the model is especially strong at writing code, planning complex projects, and reading screenshots. The launch shows that the global AI race is no longer a one-country story. Moonshot AI says its annual recurring revenue has passed 300 million dollars, and that the new model is already serving enterprise customers in more than two hundred countries. With a 27-billion-dollar valuation and an aggressive price of 15 dollars per million output tokens, K3 is putting direct pressure on the world's most expensive closed systems.