Moonshot just killed its own flagship, and almost nobody noticed the funeral. on august 31, the company quietly retired kimi-k2.5 and the entire moonshot-v1 series. api calls to those models now return a straight 404. tencent cloud pulled k2.5 the same day and told everyone still running it to migrate immediately. k2.5 lasted about eight months. released in january as moonshot's own flagship at the time, sota on agent tasks, image, and video. now it's just gone. what replaced it isn't a minor version bump. kimi k3 is 2.8 trillion parameters, a hybrid attention mechanism moonshot calls kda plus attention residuals, and a full 1 million token context window. it only activates 16 of 896 experts per token, so real compute cost stays far below what the size suggests. open weights have been public since july 27. worth being honest about where it actually sits, though. on the independent artificial analysis intelligence index, k3 scores 57.1, close behind gpt-5.6 sol max and claude fable 5, and one developer testing it on launch day called it bluntly "materially worse than sol and fable 5 for non-coding use cases." it's a specialist. strongest on coding and long-context agent work. weaker outside that lane. it even already lost the frontend crown it briefly held, kimi k2.6, a separate model still supported today, jumped from #18 to #1 on the frontend code arena within hours of its own launch, before k3 existed at all. retiring the old generation just makes k3 official by elimination. whether that's the right model for what you're actually building is a completely separate question.(CyrilXBT)