Kimi K3 scales to 2.8 T, cuts inference cost and adds multimodal support
The 2.8‑trillion‑parameter Kimi K3 expands the 48‑billion Kimi Linear model, swapping standard MoE for a LatentMoE and introducing attention residuals to boost training and validation performance. These tweaks lower inference overhead while native multimodal support opens new application paths, making K3 the largest open‑weight model yet efficient enough for broader use.