Hacker News · item history
DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf]
157
first seen points
793
peak points
793
latest points
38
observations
Posted on X
- DeepSeek's 60–85% inference speedup is the quiet shift that matters: if you can run the same model at half the latency cost, the unit economics of every AI product just changed. Not the model capability — the margin structure underneath it.