LLM Hardware Tool VRAM
VRAM
Calculator.
Select a language model, quantization, and context length to calculate exactly how much VRAM your GPU needs.
Required VRAM
5.19GB
Breakdown
Model weights4.19 GB
KV Cache (ctx 4,096)0.50 GB
Runtime overhead0.50 GB
Total estimated5.19 GB
Recommended GPU
check_circle
Minimum recommended: RTX 4060
Estimation based on Ollama / llama.cpp. Results may vary by framework.
article
New to local LLMs?
Read our full guide on how to run models on your PC with Ollama.
Model
DeepSeek-R1 7Bctx 131K
7Bparams
DeepSeek-R1 8Bctx 131K
8Bparams
DeepSeek-R1 14Bctx 131K
14Bparams
DeepSeek-R1 32Bctx 131K
32Bparams
DeepSeek-R1 70Bctx 131K
70Bparams
Gemma 2 2Bctx 8K
2.6Bparams
Gemma 2 9Bctx 8K
9.2Bparams
Gemma 2 27Bctx 8K
27Bparams
Llama 3.2 1Bctx 131K
1Bparams
Llama 3.2 3Bctx 131K
3Bparams
Llama 3.1 8Bctx 131K
8Bparams
Llama 3.1 70Bctx 131K
70Bparams
Llama 3.1 405Bctx 131K
405Bparams
Mistral 7B v0.3ctx 33K
7Bparams
Mistral Nemo 12Bctx 128K
12Bparams
Mixtral 8x7Bctx 33K
46.7Bparams
Phi-3 Mini 3.8Bctx 128K
3.8Bparams
Phi-4 14Bctx 16K
14Bparams
Phi-3 Medium 14Bctx 128K
14Bparams
Qwen2.5 7Bctx 131K
7Bparams
Qwen2.5 14Bctx 131K
14Bparams
Qwen2.5 32Bctx 131K
32Bparams
Qwen2.5 72Bctx 131K
72Bparams
Quantization
F32
starstarstarstarstar
Maximum quality. Research only.F16
starstarstarstarstar
No noticeable loss. Fine-tuning and inference.Q8
starstarstarstarstar_border
Near identical to F16, half the VRAM.Q6
starstarstarstarstar_border
Minimal loss. Good balance.Q5
starstarstarstar_borderstar_border
Recommended if you have enough VRAM.PopularQ4
starstarstarstar_borderstar_border
Most used. Best VRAM/quality balance.Q3
starstarstar_borderstar_borderstar_border
Noticeable loss. Use only if necessary.Q2
starstar_borderstar_borderstar_borderstar_border
Severe loss. Last resort.SelectedQ4_K_M — Popular
Bits/weight4.5
QualityAcceptable