DeepSeek-V4.1-Flash: open weights, vision, and a 1M context window
DeepSeek’s model card now supplies architecture details beyond the basic repository listing: image understanding, extended context, and sparse parameter activation.
Language & visionDownloadable
By Privaj Scout · Brief published Sep 11, 2026 · Sources checked Sep 16, 2026
SCOUT’S TAKE
Downloadable does not mean small. The active-parameter figure is only part of this model’s hardware story.
Inputs → output
Text + images → text
Backbone parameters
552B · not the complete model total
Conditional memory
196B additional parameters
Active parameters
8B prefill · 16B decode
Context
1M tokens
Weight format
FP8
What changed.
01
The backbone has 552B parameters, plus 196B in a separate conditional-memory component.
02
The architecture activates 8B parameters during prompt processing and 16B during generation.
Do not size a GPU from the 8B or 16B active count. Runtime, weight storage, and context-cache requirements need a measured deployment configuration; Scout has no verified minimum VRAM figure.
What the evidence shows
Architecture and license checked against DeepSeek’s model card on September 16. The displayed date remains repository creation, not a verified launch date.
ut