DeepSeek officially launched its V4.1-Flash model on September 10, 2026, marking the smallest entry in a new asymmetric MoE architecture family with native multimodal vision support. The 552B-parameter model uses a Causal Encoder-Decoder design activating just 8B parameters on input and 16B on output, delivering faster inference, higher throughput, reduced KV cache demands, and benchmark gains over the prior V4-Pro across coding, agentic tasks, and multimodal evaluations. Early beta testers highlighted significant speed improvements and lower costs, prompting DeepSeek to route V4-Pro API traffic to the new model at Flash pricing starting September 14 until V4.1-Pro arrives. This release aligns with the company’s IPO preparations and positions Flash as a competitive efficiency leader in the open-source AI landscape.