Kimi-K2-Thinking requirements — can you run it?
~1026.6B parameters · 27 quantizations measured · updated 2026-01-27 · source: unsloth/Kimi-K2-Thinking-GGUF
Every quantization: real file size vs what you actually need
| Quant | File (measured) | VRAM/RAM needed (8K ctx) | Runs on (first fit) |
|---|---|---|---|
| UD-TQ1_0 | 246.8 GB | 297.7 GB | Mac 512GB unified |
| UD-IQ1_S | 285.3 GB | 343.9 GB | Mac 512GB unified |
| UD-IQ1_M | 309.3 GB | 372.6 GB | Mac 512GB unified |
| UD-IQ2_XXS | 335.2 GB | 403.8 GB | Mac 512GB unified |
| UD-IQ2_M | 353.6 GB | 425.8 GB | Mac 512GB unified |
| Q2_K | 374.1 GB | 450.4 GB | Mac 512GB unified |
| Q2_K_L | 374.4 GB | 450.8 GB | Mac 512GB unified |
| UD-Q2_K_XL | 386.3 GB | 465.1 GB | Mac 512GB unified |
| UD-IQ3_XXS | 421.7 GB | 507.5 GB | 1TB server RAM (CPU) |
| Q3_K_S | 443.2 GB | 533.3 GB | 1TB server RAM (CPU) |
| UD-Q3_K_XL | 455.1 GB | 547.6 GB | 1TB server RAM (CPU) |
| Q3_K_M | 490.0 GB | 589.5 GB | 1TB server RAM (CPU) |
| IQ4_XS | 547.2 GB | 658.1 GB | 1TB server RAM (CPU) |
| IQ4_NL | 579.1 GB | 696.4 GB | 1TB server RAM (CPU) |
| Q4_0 | 581.2 GB | 698.9 GB | 1TB server RAM (CPU) |
| Q4_K_S | 583.3 GB | 701.5 GB | 1TB server RAM (CPU) |
| Q4_K_M | 621.2 GB | 747.0 GB | 1TB server RAM (CPU) |
| Q4_1 | 642.9 GB | 773.0 GB | 1TB server RAM (CPU) |
| UD-Q4_K_XL | 646.2 GB | 777.0 GB | 1TB server RAM (CPU) |
| Q5_K_S | 707.0 GB | 849.9 GB | 1TB server RAM (CPU) |
| Q5_K_M | 728.7 GB | 876.0 GB | 1TB server RAM (CPU) |
| UD-Q5_K_XL | 729.8 GB | 877.3 GB | 1TB server RAM (CPU) |
| Q6_K | 842.9 GB | 1.01 TB | 2TB server RAM (CPU) |
| UD-Q6_K_XL | 879.1 GB | 1.06 TB | 2TB server RAM (CPU) |
| Q8_0 | 1.09 TB | 1.31 TB | 2TB server RAM (CPU) |
| UD-Q8_K_XL | 1.19 TB | 1.43 TB | 2TB server RAM (CPU) |
| BF16 | 2.05 TB | 2.47 TB | server-grade |
KV cache uses a flat +20% context/overhead rule because this repo does not publish its config; exact numbers may differ for long contexts. MoE models keep all experts in memory — total size counts, not just active params.
Can I run Kimi-K2-Thinking on my GPU? (UD-IQ1_M recommended quant)
| Hardware | Memory | Verdict |
|---|---|---|
| RTX 3060 12GB | 12 GB | no |
| RTX 4060 Ti 16GB | 16 GB | no |
| RTX 3090 24GB | 24 GB | no |
| RTX 4090 24GB | 24 GB | no |
| RTX 5090 32GB | 32 GB | no |
| RTX PRO 6000 96GB | 96 GB | no |
| Mac 16GB unified | 16 GB | no |
| Mac 32GB unified | 32 GB | no |
| Mac 64GB unified | 64 GB | no |
| Mac 128GB unified | 128 GB | no |
| Mac 256GB unified | 256 GB | no |
| Mac 512GB unified | 512 GB | RUNS |
| 32GB system RAM (CPU) | 32 GB | no |
| 64GB system RAM (CPU) | 64 GB | no |
| 128GB system RAM (CPU) | 128 GB | no |
| 256GB system RAM (CPU) | 256 GB | no |
| 1TB server RAM (CPU) | 1024 GB | RUNS |
| 2TB server RAM (CPU) | 2048 GB | RUNS |
How much VRAM does Kimi-K2-Thinking need?
The honest answer: 372.6 GB at the UD-IQ1_M quant with 8K context — that is the measured file size (309.3 GB) plus KV cache and runtime overhead. The smallest published build needs 297.7 GB; the lossless one needs 2.47 TB.
Can I run it with CPU offload?
Yes, if combined RAM+VRAM ≥ file size — but generation speed is limited by memory bandwidth. Expect roughly: DDR5 dual channel ~10-30 tok/s for small MoE actives, single digits for big ones, 0.1-1 tok/s when experts page from disk. CPU-only is fine for batch jobs, painful for chat.
Why do other sites show different numbers?
Most pages compute weights from a formula (params × bits/8). We use the actual file sizes published in GGUF repos, which include the embedding table, unquantized tensors, and container overhead — that is why our numbers can differ from a naive calculation.
FAQ
How much VRAM does Kimi-K2-Thinking need?
The measured answer: 372.6 GB at the UD-IQ1_M quant with 8K context. The smallest published build needs 297.7 GB; the lossless (F16/BF16) build needs 2.47 TB. These are real GGUF file sizes from Hugging Face, not formula estimates.
Can I run Kimi-K2-Thinking on a 24GB GPU (RTX 3090/4090)?
Not at UD-IQ1_M (372.6 GB). Nothing fits 24GB — the smallest build needs 297.7 GB. Use the hosted API or a smaller model.
Can I run Kimi-K2-Thinking on CPU with system RAM?
Yes, if your usable RAM is at least the file size. At UD-IQ1_M you need ~372.6 GB of RAM. CPU generation is memory-bandwidth-bound: expect single-digit tok/s for large MoE models, 10-30 tok/s for small active params, and below 1 tok/s when experts page from disk.
Why do VRAM numbers for Kimi-K2-Thinking differ between sites?
Most sites compute weights from a formula (params x bits/8). ModelFit uses the actual GGUF file sizes published on Hugging Face, which include the embedding table, unquantized tensors, and container overhead — so our numbers reflect what you really download and load.