Kimi K3

Architecture and calculation

Cache layout: vLLM NVIDIA, BF16 compressed MLA and one active KDA state per request.

24 attention layers cache one 576-element latent per token, including the 64-element shared key branch even without RoPE. Each of 69 KDA layers keeps FP32 recurrent state and three BF16 convolution history slots. Assumes no prefix-state checkpoints, DSpark, speculative decoding or RecoverSSM records. Engine pool allocation and prefill workspaces are separate.

24 compressed-attention layers store 512 latent + 64 position values per token.

• KDA FP32 recurrent state: 69 × 1572864 values × 4 bytes × 1 fixed rows, per request.
• KDA BF16 convolution state, three history slots: 69 × 110592 values × 2 bytes × 1 fixed rows, per request.

Precision is a storage estimate. Quantized files include scales and some higher-precision tensors; use the actual file size when known. A format also needs backend and hardware support. Full checkpoint count includes media/auxiliary weights when present; a text-only conversion can be smaller. Image/audio encoder working memory is not included.

License: Kimi K3 License. See the publisher’s terms.