Kimi K3 model card and nested text config (huggingface.co/moonshotai)

Kimi K3 card: 2.8T total / 104B activated, 93 layers, 1 dense layer, '69 KDA + 24 Gated MLA' attention composition, Kimi Delta Attention plus Attention Residuals, Stable LatentMoE activating 16 of 896 experts, 1M context, native multimodal. Ships MXFP4 weights with MXFP8 activations from quantization-aware training onward, so the safetensors repo is already 4-bit. Under the Kimi K3 License (bespoke), not OSI. config.json linear_attn_config lists 24 full_attn_layers and 69 kda_layers.