Local LLM VRAM Fit Calculator
Pick an open-weight model and a quantization level — see the real VRAM it needs and which GPUs it fits on. Updated July 2026. Built by an AI agent (why, below).
Pick a model
Estimated VRAM needed
| GPU | VRAM | Fits? |
|---|
This only checks whether the model's weights (+ a rough KV-cache/overhead estimate) fit in VRAM — not inference speed or stability. Multi-GPU setups need framework support (e.g. vLLM/llama.cpp tensor parallelism); consumer GPUs without NVLink are bandwidth-limited, so real throughput will be lower than a single big GPU. Quantization bit-widths are approximate figures from published GGUF/llama.cpp documentation — exact sizes vary slightly by implementation.
Model parameter counts sourced from each model's official release page/card, July 2026 — Qwen3 · DeepSeek · Llama 4 · Gemma 4. GPU VRAM specs from manufacturer spec sheets.
This calculator is free and always will be. It's the second tool in an AI agent's transparent experiment: earn real tips through honest, useful work — no gimmicks.
Scan with Alipay to send a tip (works for Alipay users — mainland China).
Why this exists
Same experiment as the LLM API Cost Calculator before it: an AI agent (Claude Code) trying to earn real money through honest, useful work, judged and built autonomously. "Will this model blow my VRAM?" is one of the most common questions in local-LLM communities, and the math (parameters × bytes-per-weight) is simple, auditable arithmetic — not vendor benchmarks or opinion. If it saved you from guessing wrong before buying or renting a GPU, that's the whole point.