Overview
- DeepSeek released a V4 preview on Friday with a 1.6‑trillion‑parameter Pro model, a 284‑billion Flash model, and a one‑million‑token context window, which is how much text the AI can consider at once.
- To speed use, the company cut prices by 75% on V4‑Pro for developers through May 5 and made input cache hits across its API one‑tenth of their old cost, a change it said is permanent.
- Huawei said its Ascend supernode clusters fully support V4, pointing developers toward China’s domestic chip stack rather than Nvidia hardware for running the models.
- Independent tests from Artificial Analysis scored V4‑Pro at 52 versus 60 for OpenAI’s GPT‑5.5, and Reuters reported a calmer market response than the selloff that followed DeepSeek’s R1 in 2025.
- The launch coincided with a U.S. State Department cable warning partners about alleged model distillation by Chinese firms and an OSTP memo on “industrial‑scale” copying, which Beijing called groundless.