DeepSeek V4 Leaks Point to April Launch -- 1T Parameters, 40% Memory Reduction

TL;DR: Spec leaks circulating in Chinese media detail DeepSeek V4 as a 1-trillion-parameter model with a 1M token context window, expected in April 2026. Engineers reportedly achieved a 40% memory reduction vs V3 and 1.8x inference speedup -- meaning organizations could run a trillion-parameter model on significantly less hardware. The Hunter Alpha speculation only amplified community excitement ahead of the official drop.

Why it matters: If the leaked specs hold, DeepSeek V4 could again disrupt the economics of frontier AI just as V3 and R1 did, putting pressure on closed models at every price tier.

Source: devFlokers