Moonshot AI
Kimi K3 (1.2T)
Moonshot AI's latest frontier MoE model
Model Summary
Family
Kimi
Version
3.0
Parameters
1200B (est.)
Parameter counts for closed models are estimates; vendors rarely publish exact sizes.
VRAM Requirements by Quantization
Memory needed to serve Kimi K3 (1.2T) for inference, including a 20% overhead for activations and KV cache.
| Precision | VRAM needed | Smallest single GPU that fits |
|---|---|---|
| INT4 (4-bit) | 670.55 GB | Multi-GPU required |
| INT8 (8-bit) | 1341.10 GB | Multi-GPU required |
| FP16 (16-bit) | 2682.21 GB | Multi-GPU required |
| FP32 (32-bit) | 5364.42 GB | Multi-GPU required |
Recommended GPU Configurations
Cheapest on-demand configurations to serve Kimi K3 (1.2T) at 8-bit (1341 GB VRAM).
11x AMD Instinct MI250X
1408 GB total VRAM · CDNA 2 · multi-node
~$27.50/h
84x NVIDIA T4
1344 GB total VRAM · Turing · multi-node
~$42.00/h
7x AMD Instinct MI300X
1344 GB total VRAM · CDNA 3
~$42.00/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 Kimi K3 (1.2T)?
With an estimated 1200B parameters, Kimi K3 (1.2T) needs roughly 1341 GB of VRAM in 8-bit (INT8), 671 GB in 4-bit, and 2682 GB in FP16, including a 20% overhead for activations and KV cache.
Which GPUs can run Kimi K3 (1.2T)?
At 8-bit quantization, the most cost-effective option is 11x AMD Instinct MI250X (1408 GB combined VRAM, around $27.50/hour on-demand). Higher-end cards like the NVIDIA B200 or AMD MI355X reduce the GPU count needed.
Can Kimi K3 (1.2T) run on a single GPU?
No. Even in 4-bit, Kimi K3 (1.2T) exceeds the memory of any single current GPU, so a multi-GPU cluster is required.
How much does it cost to serve Kimi K3 (1.2T) in the cloud?
Renting 11x Instinct MI250X costs on the order of $27.50/hour, i.e. about $20,075/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 11x Instinct MI250X by the hour instead of buying hardware.
