How DeepSeek V4 Broke AI’s Cost Curse

DeepSeek V4 (Pro and Flash) has been released alongside a 58-page technical report. The release introduces two new attention mechanisms: Compressed Sparse Attention (CSA), which compresses every 4 tokens into one KV entry with sparse retrieval, and Heavily Compressed Attention (HCA), which compresses every 128 tokens for a cheap global context view. Combined with sliding window attention, this hybrid setup enables a 1 million token context window at dramatically reduced cost — V4 Pro uses only 10% of the KV cache compared to V3.2, and V4 Flash uses only 7%. DeepSeek V4 Pro is the largest open-source model at 1.6 trillion total parameters (49B active), pre-trained on 33 trillion tokens using the Muon optimizer. Post-training uses on-policy distillation from specialist models rather than direct RL on the final model. API pricing is 75% cheaper than V3.2, with DeepSeek claiming competitors cannot match this pricing without operating at a loss. The model also introduces FP4 quantization-aware training for MoE expert weights and day-zero support for Huawei chips.