Skip to content
Hacker News front page

DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and Miles

SGLang and Miles added day-0 inference and RL support for DeepSeek-V4, covering 1.6T Pro and 284B Flash. The post cites a 1M-token context, FP4 MoE expert weights, 128-token SWA, and 4:1 or 128:1 KV compression. The key systems detail is ShadowRadix coherence across three KV pools and two compression-state pools.

Why it matters: SGLang and Miles shipped inference and RL training support on DeepSeek-V4's launch day, covering the 1.6T-parameter Pro and 284B-parameter Flash versions, and disclosed deployment details including hybrid sparse attention and FP4 expert weights.

Read the original ↗Export Markdown