DeepSeek V4 Preview Launches — SGLang + Miles Deliver First Open-Source Day-0 Verified RL Pipeline for Frontier Model

LMSYS / DeepSeek / CNBC / Bloomberg·medium signal

DeepSeek unveiled V4 Pro and V4 Flash with three architecture innovations: hybrid sparse attention (1M-token context), manifold-constrained hyper-connections (mHC), and native FP4 expert weights. LMSYS published the first open-source stack to serve AND train a frontier model on launch day via SGLang+Miles. Benchmarks: 199 tok/s on B200, 266 tok/s on H200, throughput dropping only 10% from 4K to 900K context tokens. Train-inference alignment diff ~0.02-0.03 at step 0.