DeepSeek V4.1 Flash Lands on China's National Supercomputing Network with Big Cost Cuts

The DeepSeek V4.1 Flash model is now available on China's National Supercomputing Network. Users can log in to the official website, go to the Model Services page, and quickly access the model's API.

The model uses a 552B-parameter MoE architecture with a new Causal-Encoder-Decoder structure. Its input and output are asymmetric, with only 8B activated for input and 16B for output, making it much cheaper than similar-sized models.

V4.1 Flash uses a new pretraining method and larger-scale reinforcement learning post-training. In benchmarks, it outperforms several flagship models, including DeepSeek V4 Pro. It also greatly reduces KV Cache size, cutting HBM demand to one quarter and SSD demand to one eighth compared with the previous generation, lowering costs for Agent-style tasks.