DeepSeek-V4-Flash-0731

DeepSeek-V4-Flash-0731 is a dated snapshot (checkpoint from July 31, 2026) of the official DeepSeek-V4-Flash release, superseding the earlier preview version and introducing substantially stronger agentic capabilities. It belongs to the DeepSeek-V4 series built on a Mixture-of-Experts (MoE) architecture supporting up to a 1-million-token context.

The architecture uses 256 routed experts with 6 activated per token across 43 layers; the DeepSeek-V4 series introduces a hybrid attention mechanism (Compressed Sparse Attention and Heavily Compressed Attention) described in the technical report. Weights are released under the MIT license on Hugging Face (bf16/fp8 formats).

According to the model card, DeepSeek-V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) on agentic and coding benchmarks despite a substantially smaller activated parameter count.