AI VRAM Calculators
Seven free tools for planning local AI hardware. Every number is calculated from each model's published architecture.
Which tool do I need?
- You have a model in mind: LLM VRAM Calculator.
- You have a GPU and want to know what runs: Can I Run It.
- You want to train or adapt a model: Fine-Tuning VRAM Calculator.
- You generate images or video: Image & Video Model VRAM.
- Your model loads but runs out of memory on long prompts: KV Cache Calculator.
- You are choosing which file to download: GGUF Size Calculator.
- You want to know how fast it will feel: Tokens per Second Estimator.
How the tools fit together
Running a model locally comes down to three questions. First, does it fit? Memory is the sum of the model weights, the KV cache for your context, and a small runtime overhead; if the total is larger than your GPU's memory, the model either fails to load or spills into much slower system RAM. Second, how much context can you afford? On many models the cache for a long conversation or a coding agent is as large as the model itself, which is why the KV cache calculator exists as a separate tool. Third, how fast will it be? Generating text is limited by memory bandwidth, so the same model can run five times faster on one card than another with the same amount of memory.
The LLM calculator answers all three for one model. Can I Run It flips the question around for one GPU. The fine-tuning calculator adds gradients, optimizer states and activations, which is why training needs far more memory than running the same model. The image and video calculator handles diffusion pipelines, where the text encoder, the diffusion model and the VAE can be loaded one at a time.
Same formulas everywhere
All seven tools, every model and GPU page, and every guide use one shared set of formulas, so their numbers always agree. The formulas are documented on the methodology page, the underlying data is public, and every change is logged in the changelog. Results are estimates: expect real memory use within a few hundred megabytes and speeds within about 25% for typical llama.cpp, Ollama and LM Studio setups.
Tips for accurate results
- Set the context length you will actually use, not the model maximum.
- Pick the quantization of the file you will download; Q4_K_M is the most common.
- On Macs and unified-memory PCs, remember the GPU can only use part of system memory by default.
- Leave headroom: browsers and the desktop can use 0.5–1.5 GB on the GPU that drives your display.