88Z AI, 10 September — DeepSeek has officially open-sourced DeepSeek V4.1 Flash, a Mixture of Experts model with 552 billion parameters. It features an asymmetric causal encoder-decoder architecture, with only 8 billion input activation parameters and 16 billion for output, along with native multimodal capabilities. By significantly compressing the KV cache, it reduces costs for agent scenarios. The API is now live, and Tencent's WorkBuddy, CodeBuddy, and OpenCode have fully integrated the model.
At a glance
- DeepSeek V4.1 Flash is open-sourced with native multimodal capabilities and efficient KV cache compression.
- The API is live, with full integration by Tencent's WorkBuddy, CodeBuddy, and OpenCode.
The update
DeepSeek has officially announced the open-source release of the DeepSeek V4.1 Flash model. As a Mixture of Experts model with 552 billion parameters, it utilizes a unique asymmetric causal encoder-decoder architecture. Under this structure, the model's input activation parameters are only 8 billion, while the output parameters are 16 billion. It also features native multimodal processing capabilities, allowing it to directly understand and handle various types of data. To optimize practical application efficiency, the model significantly compresses the key-value cache, thereby substantially reducing usage costs in agent scenarios. Meanwhile, the relevant application programming interfaces have gone live simultaneously. Developers can access the model by setting the model name to deepseek-flash. Currently, Tencent's WorkBuddy and CodeBuddy, as well as the OpenCode platform, have fully integrated the model, marking its rapid deployment in industrial applications.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- DeepSeekSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.