Overview
- DeepSeek released the V4 preview on Friday in Pro and Flash versions with a one‑million‑token context and full support on Huawei’s Ascend 950 systems, which Huawei says also powered part of Flash’s training.
- V4‑Pro uses a mixture‑of‑experts design with 1.6 trillion parameters (49 billion active) and, by DeepSeek’s tests, leads open models in reasoning and coding while trailing frontier closed systems by roughly three to six months.
- V4‑Flash is a lighter 284‑billion‑parameter option aimed at faster, cheaper deployments that keep similar reasoning strength on many tasks but weaker world knowledge than Pro.
- Developers can access the preview now, and OpenClaw adopted V4 Flash as its default agent model on Sunday, though DeepSeek says Pro capacity is constrained and a full release date is not set.
- DeepSeek published aggressive prices—about $3.48 per million output tokens for Pro and $0.28 for Flash—while facing U.S. allegations of industrial‑scale AI copying and reported funding talks with Tencent and Alibaba valuing the lab above $20 billion.