Kimi-K3.1-Pro (2.8T-A104B) VRAM & GPU Requirement Calculator
Interactive hardware sizing for Kimi-K3.1-Pro (2.8T-A104B) (2800B by Moonshot AI). Estimated 8K context footprint: 1410.8 GB (INT4) • 2811.1 GB (FP8) • 5611.7 GB (FP16).
Memory Requirements 1410.8 GB
Requires 18 GPUs (based on memory capacity)
1400 GB
All model weights
0.31 GB
Conversation history cache
8.74 GB
Expert model optimization
1.75 GB
Temporary computation cache
Estimated using memory bandwidth and tensor compute throughput across 18x NVIDIA H100 (80GB).
vllm serve moonshotai/Kimi-K3.1-Pro --quantization awq --tensor-parallel-size 18 --max-model-len 8192
Scenario Examples (GPU + Model + Concurrency):
Click these examples to quickly configure popular model deployment scenarios!
📋 Calculation Formula FAQ
How is GPU throughput calculated?
Why do we consider memory bandwidth limitations?
How are MoE models handled differently?
What does the system efficiency factor represent?
How does L2 cache hit rate affect GPU throughput calculations?
How does precision affect performance?
Why might my results differ from other calculators?
What are the budget-friendly GPU options for LLM deployment?
What are the enterprise-grade GPU solutions?
What are the GPU requirements for popular AI models?
What are the pro tips for GPU optimization?
Kimi-K3.1-Pro (2.8T-A104B) Architecture & Memory Profile
Kimi-K3.1-Pro (2.8T-A104B) VRAM Matrix by Precision & Context Length
Precomputed single-user memory requirements (model weights + KV cache + runtime overhead).
| Precision | 8K Context | 32K Context | 128K Context |
|---|---|---|---|
| FP16 | 5611.7 GB | 5619.7 GB | 5638.9 GB |
| FP8 | 2811.1 GB | 2817.2 GB | 2828.9 GB |
| INT8 | 2811.1 GB | 2817.2 GB | 2828.9 GB |
| INT4 | 1410.8 GB | 1415.9 GB | 1423.9 GB |
Recommended GPU Hardware for Kimi-K3.1-Pro (2.8T-A104B) (INT4)
Minimum GPU count for 8K context at batch size 1 and estimated generation throughput.
| GPU Hardware | Min GPUs | Total VRAM Pool | Est. Speed |
|---|---|---|---|
| NVIDIA RTX 4090 (24GB) | 59× | 1416 GB | 40149 tok/s |
| NVIDIA RTX 5090 (32GB) | 45× | 1440 GB | 41476 tok/s |
| Apple M4 Max (128GB) (128GB) | 12× | 1536 GB | 772 tok/s |
| NVIDIA A100 (80GB) (80GB) | 18× | 1440 GB | 7706 tok/s |
| NVIDIA H100 (80GB) | 18× | 1440 GB | 13108 tok/s |
| NVIDIA H200 (141GB) | 11× | 1551 GB | 7595 tok/s |
| Huawei Ascend 910B (64GB) | 23× | 1472 GB | 7782 tok/s |