Skip to main content
株式会社オブライト

LLM VRAM Calculator

Estimate the VRAM (GPU memory) needed to run an LLM locally, from model size, quantization, and context length — and see at a glance which GPUs and Macs can handle it. Free, no sign-up.

Pick your setup — the estimate updates instantly

Estimated VRAM required

22.6 GB

Breakdown: ~18.8GB weights (31B × Q4_K_M (4bit)) + KV cache & buffers at 16K context

GPUs that fit

  • RTX 3090 / 4090 (24GB)
  • RTX 5090 (32GB)
  • RTX A6000 (48GB)
  • A100 / H100 (80GB)

Macs that fit (unified memory)

  • Mac 32GB (M1〜M4)
  • Mac 64GB (M1〜M4 Pro/Max)
  • Mac 128GB (M3/M4 Max)
  • Mac Studio 192GB (Ultra)

Estimates only — actual usage varies ±10-20% by runtime (llama.cpp / Ollama / vLLM / MLX) and settings. Models exceeding VRAM can often still run slowly via RAM offloading. Mac unified memory is counted at ~75% of total as usable.

For per-model details, see Gemma 4 Requirements and Qwen3.5-9B Local Setup.