Kimi K3 just dropped. 2.8T params, 1M context, biggest open-weight release ever.
The real story: Fireworks routed tasks between K3 and Claude Fable 5. The hybrid beat both models solo.
Open weights land July 27, but you'll need 1.4TB+ of GPU memory to self-host. For most teams the play is API access plus per-task routing.
Full specs, benchmarks, and access options: