The smallest model in DeepSeek's new architecture family relies on an asymmetric division between input and output to compress agent costs.
On September 10, 2026, DeepSeek released the V4.1-Flash version of its model, the first version of an in-house architecture called Causal Encoder-Decoder. The network has 552 billion parameters in total, but only requires 8 billion to process what is given to it as input and 16 billion to produce its response. This asymmetric slicing keeps the token calculation very low for the size of the model, which also includes native visual understanding.
The company is especially focused on memory: the attention cache requires four times less high-bandwidth memory and eight times less SSD storage than the previous generation. This is an argument directed at agents, whose bill is largely dominated by the repeated replay of a longer context. The context window reaches a little more than a million tokens.
The new rates went into effect on September 10 at 04:00 UTC, with the maintenance of a half-price off-peak schedule. Since September 14, requests to V4-Pro are automatically routed to V4.1-Flash. Weights are published on Hugging Face.
Read the full article on the original source
Go to the source →