Back to Calculator
DeepSeek V4
Hot~1 trillion total parameter MoE model from DeepSeek with ~32B active per token. Introduces Engram conditional memory (O(1) knowledge retrieval), mHC training stability, and sparse attention. 1M token context, native multimodal (text, image, video). Optimized for Huawei Ascend chips. Apache 2.0 expected.
Model Specifications
ArchitectureVISION
Parameters1000B
Familydeepseek
VRAM (Q4)500.0GB
Mixture of ExpertsActive inference parameters: 32B.
Expected late April 2026. V4 Lite accessible on API since early April 2026.
Share this Model
Send this model's specs directly to your community.
Similar Models
Related Guides
How much VRAM do you really need?
A complete breakdown of quantization levels and VRAM overhead for running local models.
Best GPUs for Machine Learning in 2026
Comparing NVIDIA and AMD options for the best speed-to-dollar ratio.
GGUF vs EXL2 vs AWQ
Understanding local AI formats and which one to pick for your specific hardware.