DeepSeek's new model needs a quarter of the memory to hold a chat.
DeepSeek released V4.1 Flash on Thursday, its first model that reads images as well as text. The 552-billion-parameter model keeps a conversation in a quarter of the fast memory the old one needed. Requests to the old flagship move over on 14 September, at the cheaper rate.
Why it mattersLess memory per conversation means cheaper long chats, and prices usually follow. A Chinese lab reaching this level again drags the whole market's prices down.
▶ Related video: DeepSeek V4.1 Flash Is INSANELY GOOD! Fast, Cheap, Powerful! (Fully Tested)
Read in the app — free, in 9 languages
Related stories
OpenAI's new image model cut generation time roughly in half.2026-09-10Inception's model writes its whole answer at once, not word by word.
2026-09-09DeepSeek will route its top model's users to a cheaper one.
2026-09-09A tiny open Chinese model matched one four times its size.
2026-09-09iFlytek trained a 293-billion-parameter model without Nvidia chips.
2026-09-07