Artificial intelligence · September 2025
DeepSeek sparse attention
Learned sparse attention makes long contexts cheap.
DeepSeek-V3.2-Exp introduced DeepSeek Sparse Attention, which attends to a selected subset of tokens instead of all of them, and cut API prices by more than half. The full V3.2 release followed on December 1, 2025, and other open labs adopted the technique.
What it built on 2
What it led to 2
Related measurements
The dashed line marks when it first worked.
Training compute of notable AI models Source: Epoch AI, Notable AI Models (CC BY 4.0)
ML hardware FLOP/s per dollar Source: Epoch AI, ML hardware (CC BY 4.0)