DeepSeek has announced the release of DeepSeek-V4.1-Flash, the smallest entry in a new architectural family designed to optimize inference speed, throughput, and multimodal capabilities. This model utilizes an asymmetric Causal Encoder–Decoder architecture within a 552B-parameter Mixture-of-Experts (MoE) framework. Technically, the design optimizes computational efficiency by employing only 8B active parameters for input processing and 16B active parameters for output generation. A significant advancement in this iteration is the reduction of KV cache requirements, which now consumes only 1/4 of the High Bandwidth Memory (HBM) and 1/8 of the SSD storage compared to the previous generation. This compression is specifically intended to lower the costs associated with cache-hit charges in agentic workflows. DeepSeek-V4.1-Flash features native visual understanding and was developed using updated pre-training methodologies alongside large-scale Reinforcement Learning (RL) post-training. The model is currently available via the DeepSeek API under the "deepseek-flash" identifier, replacing the retired V4-Flash and V4-Flash-Vision-Exp versions. To manage demand, DeepSeek has implemented a tiered pricing model where off-peak rates are set at 50% of peak rates.