DeepSeek's official launch of DeepSeek-V4 Preview models on April 24, 2026—featuring Mixture-of-Experts architectures with 1.6T total parameters (49B active) for the Pro variant and 284B (13B active) for Flash—marks a major milestone after delays from initial February rumors tied to Huawei Ascend chip optimizations. Open-sourced on Hugging Face with 1M context windows and API availability, these models deliver benchmarks rivaling closed-source leaders like Claude Opus 4.6 in agentic coding and GPT-5.4 in math, at far lower inference costs. This bolsters China's open-source AI push amid U.S. dominance, with traders eyeing rapid integrations via platforms like SiliconFlow; full V4 rollout and third-party evaluations could further solidify competitive positioning.