About ModelVRAM
ModelVRAM answers one question quickly: will this model run on my GPUs, and how many do I need? The calculators read each model's own files, show every step of the arithmetic, and run in your browser with nothing to install and no account.
Who runs it
ModelVRAM is built and maintained by Kai Lin (林凯), a developer working on AI infrastructure: LLM inference serving, vector search engines and NPU kernels, and a contributor to open-source projects in those areas. It started as the calculator Kai Lin wanted whenever a new open model came out: an answer on release day, worked out from the model's published files rather than guessed from its parameter count.
How the numbers are made
- Model shapes come from each model's files on Hugging Face: config.json and the safetensors weight metadata, or the GGUF headers. 69 preset models carry the date they were last checked.
- Weight memory is parameters × bits per parameter ÷ 8. The KV cache follows the model's own attention layout (grouped-query, multi-head latent, sliding-window and linear-attention layers), and the total adds 0.5 GB plus 10% overhead. GB means GiB (1024³ bytes).
- Speed estimates are bound by memory bandwidth and calibrated against public llama-bench runs; fine-tuning estimates count weights, gradients, optimizer states and activations as the fine-tuning calculator lays out.
- The source and tests are public, and the numbers are an open dataset.
How they are checked
Estimates are only useful if they hold up. The accuracy page compares them with 20 public VRAM measurements (median error 0.8%) and 32 llama-bench runs (median 10%). When a quantized model ships, its file sizes are checked against the weight arithmetic. Measured a model yourself? Submit the measurement and it joins the comparison.
Editorial principles
- Independent: no model maker or GPU seller pays for a listing, a number or an order.
- One referral: the “rent a GPU” card links Vast.ai with a referral code. It is disclosed where it appears and marked as sponsored, and it never changes which GPUs or providers are listed or in what order.
- Estimates, not promises: each result says what it includes and what it leaves out.
- Corrections in the open: errors are fixed as soon as they are confirmed, and every change is in the changelog.
Open data
The estimates are an open dataset in CSV and JSON under CC BY 4.0: reuse them with the attribution “Data: ModelVRAM (modelvram.com)”. See also the models that fit 8 to 96 GB and the model and GPU report.
Contact
Found a wrong number, a bug, or want a tool that does not exist yet? Write to
github.com/