DeepSeek-V4.1-Flash
DeepSeekA vision encoder (DeepSeek-ViT, trained from scratch with 2D-RoPE and 3×3 pixel-unshuffle downsampling)
1 more from this model
we replace standard absolute positional embeddings with 2D-RoPE
specific method · filed under model architecture
A rotary positional encoding adapted to two-dimensional spatial coordinates, used in vision representations.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See positional encoding for the methods filed beside it.
Also called 2D rotary position embeddings (2D-RoPE), 2D-RoPE, 2D rotary positional embeddings, Two-dimensional rotary position embeddings.
2 quotes from 2 sources, one card per model, each opening with its strongest.
A vision encoder (DeepSeek-ViT, trained from scratch with 2D-RoPE and 3×3 pixel-unshuffle downsampling)
we replace standard absolute positional embeddings with 2D-RoPE
Picked by hand, not extracted: where to read more, not evidence for anything on this page.