How to Deploy Kimi K3 Locally
Kimi K3 is a MoE model with 2,800B total parameters and 104B active per token. At Q4 it needs 1,562GB of VRAM to sit entirely on the GPU; with experts offloaded to system RAM it needs only 33GB of VRAM plus 1,531GB of RAM. No single card in the database fits it; it needs multiple GPUs or the unified memory of a Mac Studio. In offload mode the cheapest card that runs it is the CMP 170HX 40GB, at roughly 2 tokens/s.
Kimi K3 details
How much VRAM does Kimi K3 need for local deployment?
MoE: all 2,800B parameters must stay resident in memory, while speed is set by the 104B active per token.
| Quantization | WeightsGB | Total at 8KGB | Total at 32KGB | Total at 128KGB |
|---|---|---|---|---|
| 1-bit | 648.9GGUF | 658.4 | 659 | 661.5 |
| 2-bit | 861.3GGUF | 1,041.7 | 1,042.3 | 1,044.8 |
| 3-bit | 1,341.7est. | 1,342.9 | 1,343.5 | 1,346 |
| 4-bit | 1,508.7GGUF | 1,561.9 | 1,562.6 | 1,565.1 |
| 5-bit | 1,889.3est. | 1,890.5 | 1,891.1 | 1,893.7 |
| 8-bit | 1,561.2GGUF | 2,903.6 | 2,904.2 | 2,906.7 |
| FP16 | 5,476.2est. | 5,477.4 | 5,478 | 5,480.6 |
Total = weights + KV cache (FP16) + 1GB runtime overhead. Weights marked GGUF are measured file sizes from the repository; the rest are estimated from bytes per parameter. Contexts beyond this model's 1M limit show a dash. Each tier takes whichever file of that bit width the repository has, preferring unsloth dynamic quants, so the exact file name varies by model.
Which GPUs can run Kimi K3?
Xianyu
No single card in the database fits Kimi K3: it needs at least 1,562GB of VRAM at Q4_K_M with an 8K context.
A multi-GPU build needs at least 1,562GB of VRAM in total.
None of the unified-memory machines in the database fit it either.
Recommended GPUs for Kimi K3 with CPU/GPU offloading
Xianyu
Expert weights live in system RAM (needs at least 1,531 GB); VRAM holds only attention layers, shared experts and the KV cache, at least 33 GB at Q4. Speed is bound by RAM bandwidth (estimated at 70 GB/s) and is far slower than a full-VRAM setup.
| Device | VRAMGB | Price | RAM neededGB | Est. t/s |
|---|---|---|---|---|
 CMP 170HX 40GB | 40 | $1,680 | 1,531 | 1.7 |
| CMP 170HX 64GB | 64 | $2,089 | 1,531 | 1.7 |
 Instinct MI210 | 64 | $2,751 | 1,531 | 1.6 |
| Radeon PRO W7900 | 48 | $3,271 | 1,531 | 1.5 |
| GeForce RTX 4090 48GB | 48 | $3,955 | 1,531 | 1.6 |
| A100 40GB PCIe | 40 | $4,602 | 1,531 | 1.7 |
| Mac Studio M4 Max 128GB | 128 | $4,788 | 1,531 | 1.5 |
| DGX Spark 128GB | 128 | $4,832 | 1,531 | 1.4 |
| MacBook Pro M4 Max 128GB | 128 | $5,264 | 1,531 | 1.5 |
| Mac Studio M5 Max 128GB | 128 | $5,881 | 1,531 | 1.5 |
| RTX PRO 5000 Blackwell 48GB | 48 | $6,959 | 1,531 | 1.6 |
| MacBook Pro M5 Max 128GB | 128 | $8,148 | 1,531 | 1.5 |
| RTX PRO 5000 Blackwell 72GB | 72 | $9,665 | 1,531 | 1.6 |
| RTX PRO 6000D | 84 | $10,111 | 1,531 | 1.7 |
| RTX PRO 6000 Blackwell | 96 | $18,422 | 1,531 | 1.7 |
| Mac Studio M3 Ultra 512GB | 512 | $19,315 | 1,531 | 1.6 |
| Ryzen AI Max+ 395 128GB | 128 | — | 1,531 | 1.2 |
Deploy Kimi K3 locally with Ollama, llama.cpp or vLLM
The Ollama library has no official tag for it yet.
llama.cpp pulls the GGUF straight from Hugging Face; -c sets the context length.
llama-server -hf unsloth/Kimi-K3-GGUF:Q4_K_M -c 8192
vLLM serves the original-precision weights, which needs far more memory than GGUF and usually more than one GPU.
vllm serve moonshotai/Kimi-K3
File names follow that week's model-library snapshot; the definitions are in the methodology. Methodology
FAQ
How much VRAM does Kimi K3 need?
At Q4_K_M with an 8K context Kimi K3 needs about 1,562GB of VRAM; 1,876GB leaves comfortable headroom.
What is the cheapest GPU that runs Kimi K3?
No single card in the database fits Kimi K3; it needs multiple GPUs or a Mac Studio's unified memory.
Can you run Kimi K3 with Ollama?
Not yet: the Ollama library has no official tag for it. You can load the GGUF with llama.cpp instead.