DeepSeek V4 Preview Drops — 1.6T MoE Open-Source Model with 1M Context, 73% FLOP Reduction

The Verge·high signal

DeepSeek released V4 Preview on April 24 with two open-weight variants: V4-Pro (1.6T parameters, 49B activated) and V4-Flash (284B parameters, 13B activated), both supporting 1M-token context. V4 achieves 27% of V3.2's FLOPs and only 10% KV cache through a Hybrid Attention Architecture, making it dramatically more efficient. Available on Hugging Face, DeepSeek API, and chat.deepseek.com — the open-source release directly challenges GPT-5.5's launch timing.