On September 10, 2026, DeepSeek released DeepSeek-V4.1-Flash, describing it as the smallest model in a new architecture family and its first with native visual understanding. It is a 552-billion-parameter mixture-of-experts model built on a new Causal Encoder-Decoder design that activates only 8B parameters for input and 16B for output. DeepSeek says new pretraining methods plus larger-scale reinforcement-learning post-training push V4.1-Flash ahead of its flagship DeepSeek-V4-Pro on benchmarks, cost, speed and total runtime. The company also says the model's KV cache needs one-quarter the HBM and one-eighth the SSD storage of the previous generation, which it says cuts the cache-hit charges that dominate agent workloads. V4.1-Flash is now live on the DeepSeek API as the model 'deepseek-flash' with native multimodal support; the older V4-Flash and V4-Flash-Vision-Exp are retired and temporarily route to it. From 04:00 UTC on September 14, 2026, all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates until V4.1-Pro launches. New pricing took effect at 04:00 UTC on September 10, with off-peak rates at half of peak. Official partners WorkBuddy (including CodeBuddy) and OpenCode support the model.
Our takeThe headline is the retirement, not the benchmark. A lab is telling paying customers its cheaper model is the better product and routing its premium tier into it - a rare admission that the cost curve, not raw capability, is now the real frontier. For anyone running agents, cache size decides the bill more than parameter count does. Watch whether rivals answer with price cuts or a genuinely new tier.