VRAM Calculator
Estimate the GPU VRAM a workload actually needs — a local LLM with its KV cache, or an image model from Stable Diffusion 1.5 up to Flux.2. Weight sizes are the released files, not estimates from parameter counts. The result is then matched against real cards from our comparison data.
Prices and specs last checked September 15, 2026.
Configure your workload
Image models size differently: there is no KV cache, and resolution drives the overhead instead of context length.
Released weight files, checked 2026-09-13. Flux.1 dev and schnell are the same size — schnell is distilled for fewer steps, not for less memory. SD 1.5 and SDXL ship one all-in-one checkpoint; SD 3.5 and Flux ship the transformer alone, with text encoders loaded on top.
Covers the VAE, text encoders and working buffers, which run 2-3 GB above the checkpoint at 1024px and grow with resolution.
Each active control module is roughly 2 GB of extra resident weights.
Parameter count. For MoE models, we use the total parameter count (VRAM must hold all weights).
Lower quantization = smaller VRAM footprint but quality tradeoff. Q4_K_M is the most common default.
KV cache grows linearly with context. Doubling context roughly doubles the cache size.
Q8 KV cache halves memory with minor quality impact. Supported in llama.cpp and newer Ollama.
CUDA context + framework overhead. Ollama and llama.cpp are leanest.
Total VRAM needed
16.4 GB
Recommended VRAM tier
16GB card
A 16GB GPU will handle this workload with ~4GB headroom for longer context or batching.
GPUs that fit this workload
How the math works
The total VRAM needed for running an LLM is the sum of three components: model weights, KV cache, and runtime overhead. Each scales differently based on your configuration.
1. Model weights
Weight memory scales with parameter count and quantization bits per parameter. For a 7B model at Q4_K_M (~4.5 bits/param), the weights occupy roughly 4 GB. At FP16 the same model takes ~14 GB. See our how much VRAM for AI guide for a full breakdown by workload type.
weights_GB = (params_in_billions × bits_per_param) / 8
2. KV cache
Every token in your context window adds to the KV cache. What sets its size is not the model's hidden width but how many key/value heads it keeps: almost every model since 2024 uses grouped-query attention, sharing one set of keys and values across several query heads. Llama 3 70B carries 8 KV heads of 128 dimensions against 64 query heads, so its cache is about 0.31 MB per token at FP16 — roughly 10 GB at 32K context, not the 80 GB you would get by assuming every query head kept its own. The older multi-head models are the expensive ones: Llama 2 13B keeps 40 KV heads and costs 0.78 MB per token, more than twice a model five times its size.
kv_GB = 2 × num_layers × (kv_heads × head_dim) × context × precision_bytes / 1024³
3. Runtime overhead
CUDA context, framework buffers, activation memory during inference. Ollama and llama.cpp are leanest (~1-1.5 GB). vLLM with heavy batching can consume 2-3 GB more. Text Generation WebUI with extensions loaded is heaviest.
Why "recommended" differs from "fits"
A workload that needs 15.8 GB will technically fit in a 16GB card, but there's no headroom. Any extra context, a second user, or a driver update can push it over and cause OOM errors. We recommend the next VRAM tier up when headroom is under 2 GB. This is why we usually point readers toward the RTX 4090 for AI or a used RTX 3090 once workloads push past 16 GB.
This calculator gives representative estimates based on typical Transformer architectures. Actual VRAM usage varies by model family (Llama, Mistral, Qwen have slightly different layer counts and hidden sizes), driver, framework, and batch size. Treat results as ±10% accurate. See our methodology for how we derive these numbers.