DeepSeek-V4-Flash
A lighter DeepSeek-V4 serving profile that keeps the same conceptual stack while prioritizing faster deployment paths.
DeepSeek-V4-Flash matters because it shows the same paper can produce more than one practical model surface. It keeps the V4 architecture story, but the reader should expect a faster, smaller serving profile instead of the heaviest reasoning mode.
At a glance
- Family
- deepseek
- Source type
- Research
- Released
- June 2026
- Authors
- DeepSeek-AI
- Modalities
- Text
- Parameter count
- 284 billion total parameters
- Context length
- 1,048,576 tokens
- Precision
- fp8, fp4
What It Is
DeepSeek-V4-Flash is the lighter member of the V4 pair. It still belongs to the same architecture and long-context family, but it is framed as the more deployment-friendly surface.Inputs And Outputs
Like Pro, Flash is a text-in, text-out model. The distinction is operational. It aims to keep the V4 design ideas while making the serving path easier to sustain.Architecture
The same named V4 modules remain central here: compressed sparse attention, heavily compressed attention, DeepSeekMoE routing, and changed block wiring. Flash should be read as a smaller operating point on the same overall design.Output
ProbabilitiesSoftmaxOutput
LinearTransformer architectureAttention sublayer containerSparse expert sublayer containerN×RMSNormCompressed
Attention
(CSA + HCA)RMSNormDeepSeekMoEEmbeddingRoPEInput
EmbeddingInput
TokensSoftmax to Output
ProbabilitiesOutput
Linear to SoftmaxRMSNorm to Output
LinearDeepSeekMoE to RMSNormRMSNorm to DeepSeekMoERMSNorm to RMSNormCompressed
Attention
(CSA + HCA) to RMSNormEmbedding to Compressed
Attention
(CSA + HCA)Embedding to RMSNormInput
Embedding to EmbeddingRoPE to EmbeddingInput
Tokens to Input
Embedding