DeepSeek V4.1 Flash
DeepSeek · 2026-09-10
DeepSeek V4.1 Flash is DeepSeek's September 2026 multimodal mixture-of-experts model, using a causal encoder-decoder architecture and compressed sparse attention. It natively accepts text and images, generates text, and supports numeric reasoning effort from 1 to 100; the API served it through deepseek-flash until DeepSeek moved that name to DeepSeek-V4.1-Flash-1017 on October 17, 2026. Architecture dimensions describe its 40-layer text decoder, separate from the causal encoder and vision components.