DeepSeek releases V4.1-Flash, a cheaper-to-run model, and retires its V4-Pro flagship

DeepSeek says its new 552-billion-parameter model's cache needs much less memory than the previous generation's and beats V4-Pro in tests. The company says V4-Pro requests are due to route to it from September 14 and has changed its API prices. The test claims come from DeepSeek itself.

