DeepSeek V4.1 Flash: Stronger, Faster, More Accessible

Original Source from DeepSeek WeChat Official Account: https://mp.weixin.qq.com/s/qg0NU3NNUbp1co2PdkAPAg

Today we're officially releasing the DeepSeek V4.1 Flash model. It is the smallest model in our brand-new model architecture series, with native multimodal visual understanding. The new architecture was designed with these goals in mind: a higher capability ceiling, faster inference, greater throughput, and scalability to larger-parameter models.

Asymmetric architecture: big intelligence at low cost

DeepSeek V4.1 Flash is a 552B-parameter MoE model built on a brand-new Causal-Encoder-Decoder architecture. Input and output are asymmetric: only 8B parameters are activated on the input side and 16B on the output side, making it significantly cheaper than known models of the same size. V4.1 Flash also uses a new pre-training approach and has gone through larger-scale reinforcement learning post-training. In benchmark testing, it surpasses the intelligence level of a range of flagship models, including DeepSeek V4 Pro.

https://preview.redd.it/qq5p9q5qymoh1.png?width=1080&format=png&auto=webp&s=89d64fe9cfde7c480500ea0dfd543894ab2f129b

https://preview.redd.it/2tvpysuvymoh1.png?width=1080&format=png&auto=webp&s=2b4203f5faed40f47f07dfc5079150fee800b2d8

Less cache, lower cost

The new generation of models dramatically reduces the size of the KV cache. Compared with the previous generation, HBM requirements drop to 1/4 and SSD requirements to 1/8. In agent scenarios, cache-hit charges often make up a large share of the bill, so compressing the KV cache substantially lowers the cost of agent-style tasks.

Figure: DeepSeek's continued progress in reducing context storage. Relative to the first-generation model, the KV cache has shrunk 437×.

API support

DeepSeek V4.1 Flash is now live on the DeepSeek API with native multimodal support. Simply change the model name to deepseek-flash to call the latest V4.1 Flash. The older V4 Flash and V4 Flash Vision Exp models have been retired; for compatibility, the model names deepseek-v4-flash and deepseek-v4-flash-vision-exp will temporarily be routed to V4.1 Flash.

In addition, extensive testing shows that V4.1 Flash comprehensively outperforms V4 Pro on performance, cost, speed, and total time-to-completion, so we plan to phase out the V4 Pro model in an orderly fashion. After 12:00 Beijing time on September 14, 2026, and until V4.1 Pro launches, all requests to deepseek-v4-pro will be routed to V4.1 Flash and billed at V4.1 Flash's unit price.

Tencent (WorkBuddy, CodeBuddy) and OpenCode, as official partners, have now fully integrated DeepSeek V4.1 Flash — give it a try!

API pricing adjustment

Thanks to the architectural innovations, DeepSeek V4.1 Flash can serve more users at lower cost, so we have cut V4.1 Flash's pricing accordingly. To allocate resources more sensibly, we continue to use peak/off-peak pricing, with off-peak prices at half the peak rate, and encourage users to schedule tasks around their actual usage patterns. The new prices take effect at 12:00 on September 10, 2026.

https://preview.redd.it/qkhui0kjzmoh1.png?width=1690&format=png&auto=webp&s=6a03799d7586a2b19309d598b2bcd826e138f596

Open-source release

We will fully support the open-source community in adapting inference for the new model, and will explore various ways to broaden deployment. If you have large-scale deployment needs and the corresponding resources (a 2k-GPU cluster with storage cluster), please get in touch.

• Model weights: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
• Paper: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf