DeepSeek-V4 cuts long-context costs by 90% for agent workloads
deepseek-v4 targets agentic failures with hybrid attention mechanisms csa and hca. v4-pro uses 10% of kv cache memory and 27% of inference flops compared to v3.2 at 1m token context. v4-flash drops further to 7% memory. the architecture uses fp4 indexing and fp8 storage to handle long tool-use trajectories without reprompting.
DeepSeek released V4, a model architecture optimized for long-context agentic workflows. It introduces Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) to reduce inference costs. Compared to V3.2, V4-Pro uses 27% of single-token inference FLOPs and 10% of the KV cache memory at 1M tokens. V4-Flash reduces these figures to 10% FLOPs and 7% memory. This efficiency stems from 4x and 128x compression ratios, FP4 indexing, and FP8 storage.