Tencent
Hy4 preview
Tencent model discovered on huggingface
Model Summary
Family
Hy4
Version
4
Parameters
780B (est.)
Parameter counts for closed models are estimates; vendors rarely publish exact sizes.
VRAM Requirements by Quantization
Memory needed to serve Hy4 preview for inference, including a 20% overhead for activations and KV cache.
| Precision | VRAM needed | Smallest single GPU that fits |
|---|---|---|
| INT4 (4-bit) | 435.86 GB | Multi-GPU required |
| INT8 (8-bit) | 871.72 GB | Multi-GPU required |
| FP16 (16-bit) | 1743.44 GB | Multi-GPU required |
| FP32 (32-bit) | 3486.87 GB | Multi-GPU required |
Recommended GPU Configurations
Cheapest on-demand configurations to serve Hy4 preview at 8-bit (872 GB VRAM).
7x AMD Instinct MI250X
896 GB total VRAM · CDNA 2
~$17.50/h
55x NVIDIA T4
880 GB total VRAM · Turing · multi-node
~$27.50/h
5x AMD Instinct MI300X
960 GB total VRAM · CDNA 3
~$30.00/h
Quick GPU Planning
Use the calculator pre-filled with this exact version to estimate memory, speed, and compute requirements in a few clicks.
Access Pre-filled CalculatorFrequently Asked Questions
How much VRAM do you need to run Hy4 preview?
With an estimated 780B parameters, Hy4 preview needs roughly 872 GB of VRAM in 8-bit (INT8), 436 GB in 4-bit, and 1743 GB in FP16, including a 20% overhead for activations and KV cache.
Which GPUs can run Hy4 preview?
At 8-bit quantization, the most cost-effective option is 7x AMD Instinct MI250X (896 GB combined VRAM, around $17.50/hour on-demand). Higher-end cards like the NVIDIA B200 or AMD MI355X reduce the GPU count needed.
Can Hy4 preview run on a single GPU?
No. Even in 4-bit, Hy4 preview exceeds the memory of any single current GPU, so a multi-GPU cluster is required.
How much does it cost to serve Hy4 preview in the cloud?
Renting 7x Instinct MI250X costs on the order of $17.50/hour, i.e. about $12,775/month running 24/7. Actual prices vary by provider and commitment; spot and reserved capacity can be significantly cheaper.
Deploy on a GPU cloud
Rent 7x Instinct MI250X by the hour instead of buying hardware.
