mistralai_Mistral-Small-3.2-24B-Instruct-2506 requirements — can you run it?
~24.0B parameters · 26 quantizations measured · updated 2025-12-19 · source: bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF
# llama.cpp (auto-downloads the GGUF) llama-cli -hf bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF:IQ2_XS # Ollama (pulls straight from Hugging Face) ollama run hf.co/bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF:IQ2_XS # LM Studio: search "bartowski/mistralai_Mistral-Small-3.2-24B-Instruct-2506-GGUF" in the model browser
Commands load the exact quant measured on this page.
Every quantization: real file size vs what you actually need
| Quant | File (measured) | VRAM/RAM needed (8K ctx) | Runs on (first fit) |
|---|---|---|---|
| IQ2_XXS | 6.5 GB | 9.4 GB | RTX 3060 12GB |
| IQ2_XS | 7.2 GB | 10.1 GB | RTX 3060 12GB |
| IQ2_S | 7.5 GB | 10.5 GB | RTX 3060 12GB |
| IQ2_M | 8.1 GB | 11.2 GB | RTX 4060 Ti 16GB |
| Q2_K | 8.9 GB | 12.2 GB | RTX 4060 Ti 16GB |
| IQ3_XXS | 9.3 GB | 12.6 GB | RTX 4060 Ti 16GB |
| Q2_K_L | 9.5 GB | 13.0 GB | RTX 4060 Ti 16GB |
| IQ3_XS | 9.9 GB | 13.4 GB | RTX 4060 Ti 16GB |
| Q3_K_S | 10.4 GB | 14.0 GB | RTX 4060 Ti 16GB |
| IQ3_M | 10.7 GB | 14.3 GB | RTX 4060 Ti 16GB |
| Q3_K_M | 11.5 GB | 15.3 GB | RTX 3090 24GB |
| Q3_K_L | 12.4 GB | 16.4 GB | RTX 3090 24GB |
| IQ4_XS | 12.8 GB | 16.8 GB | RTX 3090 24GB |
| Q3_K_XL | 13.0 GB | 17.1 GB | RTX 3090 24GB |
| IQ4_NL | 13.5 GB | 17.7 GB | RTX 3090 24GB |
| Q4_0 | 13.5 GB | 17.7 GB | RTX 3090 24GB |
| Q4_K_S | 13.5 GB | 17.8 GB | RTX 3090 24GB |
| Q4_K_M | 14.3 GB | 18.7 GB | RTX 3090 24GB |
| Q4_K_L | 14.8 GB | 19.3 GB | RTX 3090 24GB |
| Q4_1 | 14.9 GB | 19.3 GB | RTX 3090 24GB |
| Q5_K_S | 16.3 GB | 21.1 GB | RTX 3090 24GB |
| Q5_K_M | 16.8 GB | 21.6 GB | RTX 3090 24GB |
| Q5_K_L | 17.2 GB | 22.1 GB | RTX 5090 32GB |
| Q6_K | 19.3 GB | 24.7 GB | RTX 5090 32GB |
| Q6_K_L | 19.7 GB | 25.1 GB | RTX 5090 32GB |
| Q8_0 | 25.1 GB | 31.6 GB | RTX PRO 6000 96GB |
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. Longer context grows the KV cache — use the selector above.
Can I run mistralai_Mistral-Small-3.2-24B-Instruct-2506 on my GPU? (IQ2_XS recommended quant)
| Hardware | Memory | Verdict | Est. speed |
|---|---|---|---|
| RTX 3060 12GB | 12 GB | RUNS | ~37 tok/s |
| RTX 4060 Ti 16GB | 16 GB | RUNS | ~30 tok/s |
| RTX 3090 24GB | 24 GB | RUNS | ~97 tok/s |
| RTX 4090 24GB | 24 GB | RUNS | ~105 tok/s |
| RTX 5090 32GB | 32 GB | RUNS | ~186 tok/s |
| RTX PRO 6000 96GB | 96 GB | RUNS | ~186 tok/s |
| Mac 16GB unified | 16 GB | RUNS | ~10 tok/s |
| Mac 32GB unified | 32 GB | RUNS | ~16 tok/s |
| Mac 64GB unified | 64 GB | RUNS | ~28 tok/s |
| Mac 128GB unified | 128 GB | RUNS | ~42 tok/s |
| Mac 256GB unified | 256 GB | RUNS | ~57 tok/s |
| Mac 512GB unified | 512 GB | RUNS | ~85 tok/s |
| 32GB system RAM (CPU) | 32 GB | RUNS | ~5 tok/s |
| 64GB system RAM (CPU) | 64 GB | RUNS | ~5 tok/s |
| 128GB system RAM (CPU) | 128 GB | RUNS | ~8 tok/s |
| 256GB system RAM (CPU) | 256 GB | RUNS | ~8 tok/s |
| 1TB server RAM (CPU) | 1024 GB | RUNS | ~21 tok/s |
| 2TB server RAM (CPU) | 2048 GB | RUNS | ~21 tok/s |
Speed = decode tok/s estimated from memory bandwidth ÷ measured file size (× MoE active share). Real numbers vary ±30% by runtime and settings.
How much VRAM does mistralai_Mistral-Small-3.2-24B-Instruct-2506 need?
The honest answer: 10.1 GB at the IQ2_XS quant with 8K context — that is the measured
file size (7.2 GB) plus KV cache and runtime overhead. The smallest published build needs
9.4 GB; the lossless one needs 31.6 GB.
Cheapest hardware that runs it: Used RTX 3060 12GB (~$250, used market) — see which GPU to buy.
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 mistralai_Mistral-Small-3.2-24B-Instruct-2506 need?
The measured answer: 10.1 GB at the IQ2_XS quant with 8K context. The smallest published build needs 9.4 GB; the lossless (F16/BF16) build needs 31.6 GB. These are real GGUF file sizes from Hugging Face, not formula estimates.
Can I run mistralai_Mistral-Small-3.2-24B-Instruct-2506 on a 24GB GPU (RTX 3090/4090)?
Yes — at IQ2_XS (10.1 GB). Quants that fit 24GB: IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, Q2_K, IQ3_XXS, Q2_K_L, IQ3_XS, Q3_K_S, IQ3_M, Q3_K_M, Q3_K_L, IQ4_XS, Q3_K_XL, IQ4_NL, Q4_0, Q4_K_S, Q4_K_M, Q4_K_L, Q4_1, Q5_K_S, Q5_K_M, Q5_K_L.
How fast will mistralai_Mistral-Small-3.2-24B-Instruct-2506 run?
Decode speed is memory-bandwidth-bound. At IQ2_XS: roughly 105 tok/s on an RTX 4090, 97 tok/s on an RTX 3090, 42 tok/s on an M-series Mac with 128GB, and 5.2 tok/s on CPU with dual-channel DDR4 — estimates from measured file size, MoE active-parameter share and memory bandwidth.
Can I run mistralai_Mistral-Small-3.2-24B-Instruct-2506 on CPU with system RAM?
Yes, if your usable RAM is at least the file size. At IQ2_XS you need ~10.1 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 mistralai_Mistral-Small-3.2-24B-Instruct-2506 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.