DeepSeek V4.1 Flash 1017

DeepSeek-V4.1-Flash-1017 is a multimodal Mixture-of-Experts model with a 552B-parameter backbone, activating 8B parameters per token during prefill and 16B during decoding. It natively processes images and text and supports a 1M-token context window. Its Causal Encoder-Decoder architecture and Compressed Sparse Attention 2 improve long-context inference efficiency, reducing the global KV-cache footprint to roughly one-quarter that of DeepSeek-V4-Flash. Configurable reasoning effort lets developers balance inference cost and accuracy.⁠

The 1017 update improves reasoning, coding, agentic tasks, and visual benchmark performance over the original DeepSeek-V4.1-Flash while reducing task execution time and output length. These improvements make it well suited to multimodal agents and input-heavy workflows that require long-context understanding and efficient task execution.⁠