DeepSeek released its V4.1-Flash large language model on September 10, 2026, marking the latest update to its efficient Flash tier. The 552B-parameter mixture-of-experts model introduces a new Causal Encoder-Decoder architecture that activates just 8B parameters on input and 16B on output, adds native multimodal image understanding, and delivers stronger results on coding, agentic, and multimodal benchmarks than the prior V4-Pro at lower inference cost. The MIT-licensed release retired the previous V4-Flash variant and routes legacy API calls to the new model. DeepSeek has signaled V4.1-Pro as the next release in the family while preparing an IPO on Shanghai’s STAR Market, amid intensifying competition on price-performance and open weights. No V5 timeline has been confirmed.