Data STUDIO

May 6, 2026 · Artificial Intelligence

DeepSeek V4 (Flash & Pro) Unveils Million‑Token Context and Trillion‑Parameter Inference

The April 24, 2026 release of DeepSeek V4 introduces Hybrid Attention (CSA/HCA), Manifold‑Constrained Hyper‑Connections, and the Muon optimizer, delivering 1 M‑token context windows, up to 1.6 T parameters, competitive benchmark scores against Claude and GPT, dramatically lower inference costs, and detailed deployment guidelines that expose both performance gains and practical challenges.