Mistral · 23.6B parameters

Mistral Small 24B VRAM requirements

Mistral Small 24B has 40 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 13.4 GB, and the best quantisation that fits a 24 GB card is Q6_K.

Runs comfortably

15.5 GB of 21.8 GB · 71%
022 GB
Weights 13.4 GB
KV cache 1.3 GB
Runtime overhead 0.8 GB

Mistral Small 24B at Q4_K_M leaves 6.3 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.

Generation42.6tok/s
Prompt processing1470tok/s
Max context32Ktokens
KV per 1K tokens0GB

Every quantisation of Mistral Small 24B on a RTX 4090

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 43.9 GB 46.0 GB Won't fit — 1.43
INT8 / W8A8 8.50 23.0 GB 25.1 GB Won't fit — 7.93
Q8_0 (GGUF) 8.50 23.0 GB 25.1 GB Won't fit — 7.93
FP8 (E4M3) 8.00 22.0 GB 24.0 GB Won't fit — 9.38
Q6_K 6.56 18.0 GB 20.1 GB Fits, but tight 19K 32.2
Q5_K_M 5.67 15.6 GB 17.6 GB Runs comfortably 32K 37.0
Q5_K_S 5.52 15.1 GB 17.2 GB Runs comfortably 32K 38.0
AWQ 4-bit 4.25 13.5 GB 15.6 GB Runs comfortably 32K 42.4
GPTQ 4-bit 4.25 13.5 GB 15.6 GB Runs comfortably 32K 42.4
MXFP4 4.25 13.5 GB 15.6 GB Runs comfortably 32K 42.4
Q4_K_M 4.85 13.4 GB 15.5 GB Runs comfortably 32K 42.6
Q4_K_S 4.58 12.7 GB 14.8 GB Runs comfortably 32K 44.8
Q4_0 4.55 12.6 GB 14.7 GB Runs comfortably 32K 45.1
IQ4_XS 4.25 11.9 GB 13.9 GB Runs comfortably 32K 47.9
Q3_K_M 3.91 11.0 GB 13.1 GB Runs comfortably 32K 51.5
IQ3_M 3.70 10.4 GB 12.5 GB Runs comfortably 32K 54.0
IQ3_XXS 3.06 8.8 GB 10.9 GB Runs comfortably 32K 63.3
Q2_K 2.63 7.7 GB 9.8 GB Runs comfortably 32K 71.7
IQ2_XXS 2.06 6.2 GB 8.3 GB Runs comfortably 32K 86.8
IQ1_M 1.75 5.4 GB 7.5 GB Runs comfortably 32K 98.2

Mistral Small 24B on each GPU

Q4_K_M weights at 8K context, single card, monitor attached.

GPUVRAMGB/sVerdictMax ctxtok/s
H100 SXM 80GB 80 3350 Runs comfortably 32K 161
A100 80GB 80 2039 Runs comfortably 32K 89.0
RTX 5090 32 1792 Runs comfortably 32K 82.7
RTX 4090 24 1008 Runs comfortably 32K 42.6
RTX 3090 24 936 Runs comfortably 32K 41.3
Radeon RX 7900 XTX 24 960 Runs comfortably 32K 38.6
L40S 48 864 Runs comfortably 32K 36.6
RTX A6000 48 768 Runs comfortably 32K 34.0
Mac Studio M3 Ultra 256GB 256 819 Runs comfortably 32K 33.4
Mac Studio M4 Max 128GB 128 546 Runs comfortably 32K 24.9
RTX 5080 16 960 Won't fit — 18.7
RTX 5070 Ti 16 896 Won't fit — 18.2
RTX 4080 Super 16 736 Won't fit — 15.3
RTX 4070 Ti Super 16 672 Won't fit — 14.7
RTX 5060 Ti 16GB 16 448 Won't fit — 13.1
NVIDIA DGX Spark (GB10) 128 273 Runs comfortably 32K 12.8
Mac Mini M4 Pro 48GB 48 273 Runs comfortably 32K 12.5
Ryzen AI Max+ 395 128GB 128 256 Runs comfortably 32K 9.56
RTX 4060 Ti 16GB 16 288 Won't fit — 9.09
RTX 5070 12 672 Won't fit — 6.91
RTX 4070 Super 12 504 Won't fit — 6.00
RTX 4070 12 504 Won't fit — 6.00
RTX 3060 12GB 12 360 Won't fit — 5.85
RTX 3080 10GB 10 760 Won't fit — 5.03
Arc B580 12 456 Won't fit — 4.99

Architecture

Parameters23.6B
Layers40
Hidden size5120
Attention heads / KV heads32 / 8
Head dimension128
Vocabulary131,072
Trained context32K
KV cache per 1K tokens0 GB
Hugging Facemistralai/Mistral-Small-24B-Instruct-2501

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

Mistral 7B v0.3
7.3B · 32 layersMistral NeMo 12B
12.3B · 40 layersMistral Large 2 123B
122.6B · 88 layersMixtral 8x7B
46.7B · 12.9B active · 32 layersMixtral 8x22B
140.6B · 39B active · 56 layersCodestral 22B
22.2B · 56 layers

Direct answers

Mistral Small 24B on RTX 5090
See the verdictMistral Small 24B on RTX 4090
See the verdictMistral Small 24B on RTX 3090
See the verdictMistral Small 24B on RTX 5080
See the verdictMistral Small 24B on RTX 5070 Ti
See the verdictMistral Small 24B on RTX 5070
See the verdictMistral Small 24B on RTX 4070 Ti Super
See the verdictMistral Small 24B on RTX 4070 Super
See the verdictMistral Small 24B on RTX 4070
See the verdict