Mistral · Codestral · 22.2B parameters
Codestral 22B VRAM requirements
Codestral 22B has 56 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 12.6 GB, and the best quantisation that fits a 24 GB card is Q6_K.
Runs comfortably
15.2 GB of 21.8 GB · 70%Codestral 22B at Q4_K_M leaves 6.6 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.
Every quantisation of Codestral 22B 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 | 41.4 GB | 44.0 GB | Won't fit | — | 1.56 |
| INT8 / W8A8 | 8.50 | 21.9 GB | 24.5 GB | Won't fit | — | 9.30 |
| Q8_0 (GGUF) | 8.50 | 21.9 GB | 24.5 GB | Won't fit | — | 9.30 |
| FP8 (E4M3) | 8.00 | 20.7 GB | 23.3 GB | Won't fit | 1K | 12.0 |
| Q6_K | 6.56 | 17.0 GB | 19.6 GB | Runs comfortably | 18K | 33.5 |
| Q5_K_M | 5.67 | 14.7 GB | 17.3 GB | Runs comfortably | 29K | 38.4 |
| Q5_K_S | 5.52 | 14.3 GB | 16.9 GB | Runs comfortably | 30K | 39.4 |
| Q4_K_M | 4.85 | 12.6 GB | 15.2 GB | Runs comfortably | 32K | 44.3 |
| Q4_K_S | 4.58 | 11.9 GB | 14.5 GB | Runs comfortably | 32K | 46.6 |
| Q4_0 | 4.55 | 11.8 GB | 14.4 GB | Runs comfortably | 32K | 46.9 |
| AWQ 4-bit | 4.25 | 11.5 GB | 14.1 GB | Runs comfortably | 32K | 47.9 |
| GPTQ 4-bit | 4.25 | 11.5 GB | 14.1 GB | Runs comfortably | 32K | 47.9 |
| MXFP4 | 4.25 | 11.5 GB | 14.1 GB | Runs comfortably | 32K | 47.9 |
| IQ4_XS | 4.25 | 11.0 GB | 13.6 GB | Runs comfortably | 32K | 49.9 |
| Q3_K_M | 3.91 | 10.2 GB | 12.8 GB | Runs comfortably | 32K | 53.7 |
| IQ3_M | 3.70 | 9.6 GB | 12.3 GB | Runs comfortably | 32K | 56.4 |
| IQ3_XXS | 3.06 | 8.0 GB | 10.6 GB | Runs comfortably | 32K | 66.4 |
| Q2_K | 2.63 | 6.9 GB | 9.5 GB | Runs comfortably | 32K | 75.5 |
| IQ2_XXS | 2.06 | 5.5 GB | 8.1 GB | Runs comfortably | 32K | 92.1 |
| IQ1_M | 1.75 | 4.7 GB | 7.3 GB | Runs comfortably | 32K | 105 |
Codestral 22B on each GPU
Q4_K_M weights at 8K context, single card, monitor attached.
| GPU | VRAM | GB/s | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|
| H100 SXM 80GB | 80 | 3350 | Runs comfortably | 32K | 165 |
| A100 80GB | 80 | 2039 | Runs comfortably | 32K | 92.0 |
| RTX 5090 | 32 | 1792 | Runs comfortably | 32K | 85.6 |
| RTX 4090 | 24 | 1008 | Runs comfortably | 32K | 44.3 |
| RTX 3090 | 24 | 936 | Runs comfortably | 32K | 43.0 |
| Radeon RX 7900 XTX | 24 | 960 | Runs comfortably | 32K | 40.1 |
| L40S | 48 | 864 | Runs comfortably | 32K | 38.1 |
| RTX A6000 | 48 | 768 | Runs comfortably | 32K | 35.3 |
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 32K | 34.8 |
| Mac Studio M4 Max 128GB | 128 | 546 | Runs comfortably | 32K | 25.9 |
| RTX 5080 | 16 | 960 | Won't fit | 4K | 20.8 |
| RTX 5070 Ti | 16 | 896 | Won't fit | 4K | 20.2 |
| RTX 4080 Super | 16 | 736 | Won't fit | 4K | 16.8 |
| RTX 4070 Ti Super | 16 | 672 | Won't fit | 4K | 16.1 |
| RTX 5060 Ti 16GB | 16 | 448 | Won't fit | 4K | 14.3 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Runs comfortably | 32K | 13.4 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Runs comfortably | 32K | 13.0 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Runs comfortably | 32K | 9.97 |
| RTX 4060 Ti 16GB | 16 | 288 | Won't fit | 4K | 9.75 |
| RTX 5070 | 12 | 672 | Won't fit | — | 7.21 |
| RTX 4070 Super | 12 | 504 | Won't fit | — | 6.26 |
| RTX 4070 | 12 | 504 | Won't fit | — | 6.26 |
| RTX 3060 12GB | 12 | 360 | Won't fit | — | 6.11 |
| Arc B580 | 12 | 456 | Won't fit | — | 5.21 |
| RTX 3080 10GB | 10 | 760 | Won't fit | — | 5.16 |
Architecture
| Parameters | 22.2B |
| Layers | 56 |
| Hidden size | 6144 |
| Attention heads / KV heads | 48 / 8 |
| Head dimension | 128 |
| Vocabulary | 32,768 |
| Trained context | 32K |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | mistralai/Codestral-22B-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