Kimi K3 Open Weights Drop July 27: High Hallucination Risk

Moonshot AI plans to release the full model weights for Kimi K3 on July 27. This 2.8-trillion-parameter model recently topped the Arena.ai Frontend Code Arena leaderboard. Here's the thing: while K3 performs well in coding, independent testing shows its hallucination rate increased to approximately 51% from its predecessor's 39%. This figure represents incorrect answers divided by all non-correct responses. What's interesting is that K3 simultaneously improved at answering questions but worsened at recognizing when it's wrong. The bottom line is this trade-off matters, especially for tasks requiring factual accuracy. The July 27 release would make K3 the first downloadable model in its parameter class, allowing organizations to self-host and fine-tune it.