DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf]

**Brief Summary (2‑3 sentences)** DeepSeek has released a set of open‑source inference‑time optimizations that accelerate text generation by 60 %–85 % compared