DeepSeek V4 Launch Reportedly Imminent: 1T-Parameter MoE Delayed by Huawei Ascend Chip Rewrite

NxCode·medium signal

The Information reported on April 3 that DeepSeek V4's launch is imminent after delays caused by rewriting code for Huawei Ascend and Cambricon chips, denying NVIDIA early access. The model is a ~1 trillion parameter MoE with ~37B active parameters, 1M-token context via Engram conditional memory (97% needle-in-haystack accuracy), and native multimodal generation. A 'V4 Lite' appeared on DeepSeek's website March 9, suggesting incremental rollout. Full release expected within weeks.