DeepSeek-V4-Flash-0731

DeepSeek · DeepSeek
LLMMITpublicTrusted304B params

DeepSeek V4 Flash, 31.07.2026 update: 304B MoE (FP8 + FP4 experts).

DeepSeek-V4-Flash-0731

Precision

The checkpoint is mixed: the routed experts are FP4 (e2m1 values packed two per byte, one e8m0 scale per 32 values) - about 97 % of the parameters; attention and the shared expert are FP8 e4m3 with e8m0 block scales; norms and gates are BF16/F32. Its quantization_config declares fp8, so a task pins Fp8 and the miner loads the checkpoint as published.

Memory

The floor follows the 166.9 GB of weights on disk plus 10 % for the KV cache and activations - not BF16 arithmetic.