A leaked test interface and developer API observations suggest DeepSeek V4 is weeks away from public release as a suite of models—Lite, standard, and Vision tiers—with new Fast, Expert, and Vision interaction modes. The full release will run on Huawei Ascend 950PR chips rather than NVIDIA hardware, marking a complete shift to domestic Chinese semiconductor infrastructure. Early API testers report a 30% inference speed increase and context recall rising from 45% to 94% at 128K tokens versus V3.