Mistral · Mixtral · 140.6B parameters · 39B active
Mixtral 8x22B VRAM requirements
Mixtral 8x22B has 56 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 79.4 GB.
Won't fit
82.0 GB of 21.8 GB · 377%Short by 60.3 GB. You can run it with 13 of 56 layers on the RTX 4090 and the rest in system RAM, at roughly 2.12 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.
Every quantisation of Mixtral 8x22B on a RTX 4090
Highlighted row is the highest quality that still fits at 8K context.
| Quantisation | bpw | Weights | Total | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 261.9 GB | 264.5 GB | Won't fit | — | 0.56 |
| INT8 / W8A8 | 8.50 | 139.0 GB | 141.6 GB | Won't fit | — | 1.10 |
| Q8_0 (GGUF) | 8.50 | 139.0 GB | 141.6 GB | Won't fit | — | 1.10 |
| FP8 (E4M3) | 8.00 | 130.9 GB | 133.5 GB | Won't fit | — | 1.18 |
| Q6_K | 6.56 | 107.4 GB | 110.0 GB | Won't fit | — | 1.46 |
| Q5_K_M | 5.67 | 92.8 GB | 95.4 GB | Won't fit | — | 1.75 |
| Q5_K_S | 5.52 | 90.4 GB | 93.0 GB | Won't fit | — | 1.80 |
| Q4_K_M | 4.85 | 79.4 GB | 82.0 GB | Won't fit | — | 2.12 |
| Q4_K_S | 4.58 | 75.0 GB | 77.6 GB | Won't fit | — | 2.29 |
| Q4_0 | 4.55 | 74.5 GB | 77.1 GB | Won't fit | — | 2.31 |
| AWQ 4-bit | 4.25 | 70.1 GB | 72.7 GB | Won't fit | — | 2.50 |
| GPTQ 4-bit | 4.25 | 70.1 GB | 72.7 GB | Won't fit | — | 2.50 |
| MXFP4 | 4.25 | 70.1 GB | 72.7 GB | Won't fit | — | 2.50 |
| IQ4_XS | 4.25 | 69.6 GB | 72.2 GB | Won't fit | — | 2.51 |
| Q3_K_M | 3.91 | 64.1 GB | 66.7 GB | Won't fit | — | 2.78 |
| IQ3_M | 3.70 | 60.6 GB | 63.3 GB | Won't fit | — | 3.00 |
| IQ3_XXS | 3.06 | 50.2 GB | 52.8 GB | Won't fit | — | 3.94 |
| Q2_K | 2.63 | 43.2 GB | 45.8 GB | Won't fit | — | 4.91 |
| IQ2_XXS | 2.06 | 33.9 GB | 36.5 GB | Won't fit | — | 7.58 |
| IQ1_M | 1.75 | 28.8 GB | 31.4 GB | Won't fit | — | 11.0 |
Mixtral 8x22B on each GPU
Q4_K_M weights at 8K context, single card, monitor attached.
| GPU | VRAM | GB/s | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 64K | 17.4 |
| H100 SXM 80GB | 80 | 3350 | Won't fit | — | 15.1 |
| Mac Studio M4 Max 128GB | 128 | 546 | Runs comfortably | 64K | 13.0 |
| A100 80GB | 80 | 2039 | Won't fit | — | 12.3 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Runs comfortably | 64K | 6.72 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Runs comfortably | 64K | 5.02 |
| RTX A6000 | 48 | 768 | Won't fit | — | 3.27 |
| L40S | 48 | 864 | Won't fit | — | 3.16 |
| RTX 5090 | 32 | 1792 | Won't fit | — | 2.65 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Won't fit | — | 2.55 |
| RTX 3090 | 24 | 936 | Won't fit | — | 2.21 |
| RTX 4090 | 24 | 1008 | Won't fit | — | 2.12 |
| RTX 5080 | 16 | 960 | Won't fit | — | 2.12 |
| RTX 5070 Ti | 16 | 896 | Won't fit | — | 2.11 |
| RTX 5060 Ti 16GB | 16 | 448 | Won't fit | — | 2.09 |
| Radeon RX 7900 XTX | 24 | 960 | Won't fit | — | 2.01 |
| RTX 5070 | 12 | 672 | Won't fit | — | 2.00 |
| RTX 4080 Super | 16 | 736 | Won't fit | — | 1.91 |
| RTX 4070 Ti Super | 16 | 672 | Won't fit | — | 1.91 |
| RTX 3060 12GB | 12 | 360 | Won't fit | — | 1.87 |
| RTX 4060 Ti 16GB | 16 | 288 | Won't fit | — | 1.87 |
| RTX 3080 10GB | 10 | 760 | Won't fit | — | 1.86 |
| RTX 4070 Super | 12 | 504 | Won't fit | — | 1.81 |
| RTX 4070 | 12 | 504 | Won't fit | — | 1.81 |
| Arc B580 | 12 | 456 | Won't fit | — | 1.53 |
Architecture
| Parameters | 140.6B |
| Active per token | 39B of 8 experts, top-2 |
| Layers | 56 |
| Hidden size | 6144 |
| Attention heads / KV heads | 48 / 8 |
| Head dimension | 128 |
| Vocabulary | 32,768 |
| Trained context | 64K |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | mistralai/Mixtral-8x22B-Instruct-v0.1 |
The Mistral family
Dense models — 7B, NeMo 12B, Small 24B and Large 123B — alongside the two Mixtral mixture-of-experts releases and Codestral for code. Vocabulary size is not consistent across the family: 32k on 7B, Large and Codestral, 131k on NeMo and Small, which changes how much of a small model is embedding table.
huggingface.co/mistralai · mistral.ai · all 7 Mistral models