DeepSeek releases DeepSeek-V4.1-Flash with a causal encoder–decoder design

Background

On 10 September 2026 DeepSeek added a new model to its lineup. DeepSeek described the model as the smallest in its new architecture family, with native visual understanding ^[1].

What happened

On 10 September 2026 DeepSeek released DeepSeek-V4.1-Flash under the API model name deepseek-flash ^[1]. The model is multimodal and replaced V4-Flash and V4-Flash-Vision-Exp ^[1]. From 14 September 2026 it also served requests sent to deepseek-v4-pro, until a V4.1-Pro model launches ^[1].

Key details

The model is a 552-billion-parameter multimodal mixture-of-experts model, with a new causal encoder-decoder design. It uses 8 billion active parameters for input (prefill) and 16 billion for output (decode) ^[1][2]. The architecture consists of 40 layers, a 20-layer encoder and a 20-layer decoder, with 384 routed experts per MoE layer and 6 of them active per token ^[2].

Pre-training used 45 trillion tokens of multimodal data, and the context was extended to 1M tokens. Reasoning effort can be adjusted from 1 to 100 ^[2].

The KV cache needs a quarter of the HBM and an eighth of the SSD storage of the previous generation, at about 890 bytes per token ^[1][2].

The weights and a technical report were published on Hugging Face under the MIT License ^[1][2]. DeepSeek said its benchmark results were ahead of flagship models including DeepSeek-V4-Pro ^[1].

Impact

For developers, the lower KV-cache requirement means that long conversations need less memory than on the previous generation ^[1][2].

Terms in the release

Multimodal means that a model can work with more than one kind of input, here text and images. The model has native visual understanding ^[1].

Active parameters are the parameters used in each step of computation. In a mixture-of-experts model, only part of the network runs for a given token, so the active count is much smaller than the total ^[1][2].

The KV cache stores attention data for tokens the model has already processed, so that it does not need to recompute them as a conversation grows. Because the cache grows with the length of the conversation, reducing its size directly affects how much memory a long session needs. HBM, or high-bandwidth memory, is the fast memory attached to AI accelerators, and SSD storage is the larger, slower storage of a solid-state drive. The release compares the previous generation's use of both ^[1][2].