DeepSeek V4 preview: 1.6T MoE, 1M context, Apache 2.0 — and a Huawei Ascend pivot

DeepSeek released preview V4-Pro-Max (1.6T parameters) and V4 Flash (284B) MoE models with 1M-token context, open-weighted under Apache 2.0 at API pricing of $1.74/$3.48 per million tokens. The release introduces Compressed Sparse Attention + Heavily Compressed Attention, cutting KV cache by ~10x at 1M context and using just 27% of single-token inference FLOPs vs V3.2. Crucially, DeepSeek validated fine-grained Expert Parallelism on Huawei Ascend NPUs at 1.5–1.73x speedups — a blueprint resilient to US export controls.