DeepSeek V4 Pro 0813

Architecture and calculation

Cache layout: DeepSeek reference BF16 cache (text, DSpark disabled).

Persistent cache uses BF16 vectors and FP32 pooling state; Q8 does not shrink it. Assumes text generation with speculative decoding disabled. Long prompts require chunked prefill; the reference demo’s full-prefill peak is not included.

• Shared K/V sliding ring (BF16): 61 × 512 values × 2 bytes × 128 fixed rows, per request.
• CSA compressed shared K/V (BF16): 30 × 512 values × 2 bytes per 4 tokens, per request.
• HCA compressed shared K/V (BF16): 31 × 512 values × 2 bytes per 128 tokens, per request.
• CSA compressed index vectors (BF16): 30 × 128 values × 2 bytes per 4 tokens, per request.
• CSA pooling values and scores (FP32): 30 × 2048 values × 4 bytes × 8 fixed rows, per request.
• HCA pooling values and scores (FP32): 31 × 1024 values × 4 bytes × 128 fixed rows, per request.
• CSA index pooling values and scores (FP32): 30 × 512 values × 4 bytes × 8 fixed rows, per request.
• Shared rotary frequency tables (complex64): 1 × 32 values × 8 bytes per token, shared across requests.

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 may include optional MTP or draft weights; cached speculative tokens are extra.

License: MIT. See the publisher’s terms.