Tools / LLM GPU calculator / Mistral Small 3.2 24B
Mistral Small 3.2 24B VRAM requirements
GPU memory for the weights and KV cache at FP16, FP8 and 4-bit, and how many GPUs that takes. Figures from the model's config.json.
Short answer
Mistral Small 3.2 24B needs about 55 GB of GPU memory at FP16/BF16, 28 GB at FP8 and 16 GB at 4-bit to serve one user at an 8K-token context. Its 24B parameters take 48 GB at 16-bit, and that conversation adds 1.3 GB of KV cache. On NVIDIA H100 80 GB GPUs that is 1 GPU at FP16 or FP8; for a desktop, at 4-bit it fits a single 24 GB GeForce RTX 4090.
Updated . KV cache at FP16, plus 10% headroom for activations and the runtime.
GPU memory by precision
| Weights | Weight memory | 1 user, 8K context | 32 users, 8K context | 1 user, 128K context |
|---|---|---|---|---|
| FP16/BF16 | 48 GB | 55 GB | 101 GB | 77 GB |
| FP8 | 24 GB | 28 GB | 74 GB | 51 GB |
| 4-bit (INT4) | 13.2 GB | 16 GB | 62 GB | 39 GB |
Required memory = (weights + KV cache) ÷ 0.9. An FP8 KV cache halves the cache share; set it in the calculator.
How many GPUs for one user at 8K context
| GPU | Memory | FP16/BF16 | FP8 | 4-bit (INT4) |
|---|---|---|---|---|
| Data center · tensor parallel | ||||
| NVIDIA L40S | 48 GB | 2 GPUs · TP2 | 1 GPU | 1 GPU |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 96 GB | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H100 SXM | 80 GB | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H100 NVL | 94 GB | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H200 | 141 GB | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B200 | 180 GB | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B300 | 288 GB | 1 GPU | 1 GPU | 1 GPU |
| AMD Instinct MI300X | 192 GB | 1 GPU | 1 GPU | 1 GPU |
| AMD Instinct MI355X | 288 GB | 1 GPU | 1 GPU | 1 GPU |
| Workstation & consumer · layer split | ||||
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | 1 card | 1 card | 1 card |
| NVIDIA GeForce RTX 3090 | 24 GB | 3 cards | 2 cards | 1 card |
| NVIDIA GeForce RTX 4090 | 24 GB | 3 cards | 2 cards | 1 card |
| NVIDIA GeForce RTX 5090 | 32 GB | 2 cards | 1 card | 1 card |
| Apple Mac Studio M3 Ultra / M5 Ultra | 96 GB (72 usable) | 1 Mac | 1 Mac | 1 Mac |
| Apple Mac Studio M3 Ultra / M5 Ultra | 256 GB (192 usable) | 1 Mac | 1 Mac | 1 Mac |
| Apple Mac Studio M5 Ultra | 512 GB (384 usable) | 1 Mac | 1 Mac | 1 Mac |
Data center GPUs are counted at a tensor-parallel size (TP) that divides the model’s attention heads, within one 8-GPU server. Workstation and consumer cards are counted for a layer split, which works with any number of cards. Mac Studio counts about 75% of unified memory as usable by the GPU. Capacity only: speed depends on memory bandwidth and the serving stack.
Memory is step one. Throughput, latency, power and the network decide the real build.
Size my actual workloadMistral Small 3.2 24B architecture
- Parameters
- 24B
- Layers with a KV cache
- 40
- Attention heads / KV heads
- 32 / 8
- Head dimension
- 128
- Attention
- Grouped-query attention (4 query heads per KV head)
- Maximum context
- 131,072 tokens
- KV cache per token (FP16)
- 2 × 40 layers × 8 KV heads × 128 head dim × 2 bytes = 163,840 bytes (0.16 MB) per token
Source: mistralai/Mistral-Small-3.2-24B-Instruct-2506 config.json on Hugging Face. Parameter counts are the Hub’s count of the repo’s safetensors files, or the model card’s figure.
Questions about Mistral Small 3.2 24B
How much VRAM does Mistral Small 3.2 24B need?
About 55 GB at FP16/BF16, 28 GB at FP8 and 16 GB at 4-bit, for one user at an 8K-token context. That is 48 GB, 24 GB or 13.2 GB of weights, plus 1.3 GB of KV cache, plus 10% headroom for activations and the runtime. Serving 32 users at 8K each takes 101 GB at FP16 or 74 GB at FP8, because every conversation keeps its own KV cache.
How many H100 GPUs does Mistral Small 3.2 24B need?
At FP16/BF16 it takes 1 GPU for one user at 8K and 2 GPUs (tensor parallel 2) for 32 users. At FP8, 1 GPU, for one user at 8K or for 32 users. One user at FP8 takes 1 GPU on H200 (141 GB) and 1 GPU on B200 (180 GB).
Can Mistral Small 3.2 24B run on an RTX 4090 or RTX 5090?
At 4-bit it needs about 16 GB for one user at 8K: a single RTX 4090 (24 GB) or a single RTX 5090 (32 GB). At FP8 (28 GB) it takes 2 RTX 4090s or a single RTX 5090 (32 GB). Several cards work with a layer split, as in llama.cpp, which divides the model's layers between them.
Can Mistral Small 3.2 24B run on a Mac Studio?
macOS lets the GPU use about 75% of unified memory by default: about 72 GB of a 96 GB, 192 GB of a 256 GB and 384 GB of a 512 GB Mac Studio. At 4-bit Mistral Small 3.2 24B needs 16 GB for one user at 8K, so it fits one 96 GB Mac Studio; at FP8 (28 GB) it fits one 96 GB Mac Studio.
How large is Mistral Small 3.2 24B's KV cache?
At FP16, 2 × 40 layers × 8 KV heads × 128 head dim × 2 bytes = 163,840 bytes (0.16 MB) per token. It uses grouped-query attention: 4 query heads share each KV head, which makes the cache 4× smaller than full multi-head attention. An 8K-token conversation holds 1.3 GB at FP16, and a 128K-token one holds 21 GB.
What is Mistral Small 3.2 24B's maximum context length?
131,072 tokens, according to its Hugging Face configuration. One user at that full length holds 21 GB of KV cache at FP16, for 51 GB in total with FP8 weights, or 39 GB with an FP8 KV cache as well.
Other models
- Llama 3.1 8B
- Llama 3.3 70B
- Llama 3.1 405B
- Llama 4 Scout
- Llama 4 Maverick
- Qwen3 8B
- Qwen3 14B
- Qwen3 32B
- Qwen3-30B-A3B
- Qwen3-235B-A22B
- Qwen2.5 32B
- Qwen2.5 72B
- DeepSeek-V3 / R1
- gpt-oss-20b
- gpt-oss-120b
- Gemma 3 27B
- Mixtral 8x7B
- Phi-4 14B
- GLM-5.3
- LFM2.5-24B-A2B
- Kolibri-1
Not listed? The LLM GPU calculator sizes any Hugging Face model from its ID, link or config.json.
Need to deploy private AI?
We design it, validate it, source it, deploy it and help run it.
Give us your workload, users, budget and timeline. Within 48 hours we'll give you a validated cloud/on-prem architecture, BOM, expected costs and sourcing options.