DeepSeek V4 expected late April with 1 trillion parameters optimized for Huawei Ascend chips

DeepSeek V4 is anticipated in late April 2026 featuring approximately 1 trillion total parameters using Mixture-of-Experts architecture with only 32-37 billion parameters active per token. The model has missed two previous launch windows but represents the first frontier-class model optimized for Chinese domestic semiconductors after months of collaboration with Huawei and Cambricon.