DeepSeek V4.1 Flash is a 552B-parameter MoE model featuring a novel Causal-Encoder-Decoder architecture with asymmetric inputs and outputs. With an input activation of only 8B and an output activation of 16B, its cost is significantly lower than existing models of the same size.
Different token groups have different prices. Unit: Million tokens (M)
| Token Group | Description | Group Ratio | Input Price | Output Price | Cache Read |
|---|---|---|---|---|---|
| deepseek | Deepseek | 0.5x | $0.5/M | $2/M | $0.01/M |