Source Posts on X
The top open source models like Kimi K3 clearly are at par with most capabilities. The two areas it's worse at are 1) it's not as detail oriented and gets things wrong more, and 2) can't solve the very top echelon of questions, like open math problems. Both feel a data and post training type problem, which codex and claude have a headstart in because they even have user data galore, but these don't feel like they'll be enduring advantages forever. The multi agent work that the harnesses are doing will continue to make the lead stay but it's not some intrinsic capability barrier. The OS world, as a lesson, still had behemoths in Windows and macos even while Linux, an equally capable OS that had rough edges and wasn't as user friendly, so this isn't counting anyone out. But capabilities aren't monolithic or the only thing that matters to users, or for market value, it's worth remembering this.