One set of parameters processes the prompt and a larger one generates the answer, and DeepSeek says the design cuts the memory for cached context to a quarter of V4-Flash's.
On September 10 DeepSeek released V4.1-Flash, the smallest model in its new architecture family. The split is 8 billion active parameters for the input and 16 billion for the output. The model handles text and images, accepts up to a million tokens of context and comes with open weights under the MIT license.
DeepSeek puts the global context cache at 890 bytes per token. For SSD storage it reports a footprint of about an eighth of V4-Flash's, because the model reconstructs the cache for the most recent tokens when needed instead of saving that portion to disk.
Running at maximum reasoning effort, V4.1-Flash scores 40 on the Intelligence Index of Artificial Analysis, an independent benchmarking firm, above the 36 of DeepSeek's larger V4 Pro.
The firm estimates an average cost of $0.27 per benchmark task against $0.67 for V4 Pro, even though V4.1-Flash used 89,000 output tokens per task, reasoning included, to V4 Pro's 55,000. GLM-5.3-Flash does slightly better on both counts, at 42 points and $0.25.
At peak hours, DeepSeek charges $0.30 per million uncached input tokens, $1.20 for output and $0.006 for cached input. Off-peak rates are half those amounts.
DeepSeek has also dropped a plan to send all V4 Pro requests to the new model from September 14, citing user demand.