What Happened
DeepSeek launched V4.1 Flash through its API on September 10, 2026. The technical report followed later. The official model repository confirms downloadable weights under MIT; this entry uses the API launch date without assuming that every research artifact appeared simultaneously.
Why It Matters
V4.1 Flash introduced a notable departure from uniform decoder-only processing. Separating the computation used to ingest context from the computation used to generate output targets input-heavy agents, whose growing histories can make memory and prefill costs dominant.
Technical Details
The model has a 552-billion-parameter backbone, native image and text input, and context support up to one million tokens. Its Causal Encoder-Decoder uses 20 encoder and 20 decoder layers, activating eight billion parameters during prefill and 16 billion during decoding. Shared and compressed key-value state reduces cache requirements. Additional conditional-memory components mean the backbone count should not be mistaken for the entire stored checkpoint.