Kimi K3 broke 90%! - Revise

I got around to benchmarking Kimi K3 on ErrataBench and it's the first open weights model1 to break 90%! Kimi beats Opus 4.8 on both the performance and price axes, and is close to 5.6 Terra-level performance.<br>What does that mean? It means if you're running a local model and you have the hardware for it, Kimi K3 is the best choice for tasks related to writing and editing in the English language. ErrataBench measures how well models can spot errors in large samples of text and introduce valid fixes using tools.<br>In my opinion, this is a milestone. The leaderboard on this benchmark has been dominated by proprietary models - this is the first time an open model has broken into the top 10.

1 Well, a soon-to-be-open model, supposedly