Moonshot AI published the full open weights for Kimi K3 on July 26, roughly 7:30 PM EDT, a day ahead of the July 27 target it had previously communicated. The model carries 2.8 trillion total parameters, a 1,048,576-token context window, and native visual understanding, built on two internally developed architectural innovations called Kimi Delta Attention and Attention Residuals. Cloud platforms Together AI and Modal both announced day-0 hosted access timed to the release.
At 2.8 trillion parameters, K3 is roughly 75% larger than DeepSeek's V4 Pro and marks the first open-weight model to reach the 3-trillion-parameter class. The full weights total approximately 1.4 terabytes using MXFP4 four-bit quantization, meaning self-hosting requires substantial multi-GPU infrastructure; most teams will access the model through inference providers rather than running it locally. Only 16 of 896 experts fire per token, so per-token compute resembles a mid-size model despite the total parameter count.
The release follows K3's initial API and consumer-app launch on July 16, when it topped Arena.ai's Frontend Code Arena leaderboard in blind developer testing. The model ships under a Modified MIT license permissive enough for commercial use. Independent benchmarking beyond Moonshot's own reported scores remains limited, and some coverage has flagged elevated hallucination rates in early third-party testing that did not appear in the company's own benchmark charts.
