Can I run this model locally?
Pick a model from the catalog or enter its parameter count to estimate the memory it needs at FP16, 8-bit and 4-bit, plus KV-cache headroom for your context length. All figures are rough estimates.
Choose a model or enter a parameter count to see an estimate.
How the estimate works
Weights = parameters × bits per weight: 16 for FP16/BF16, about 8.5 for 8-bit and about 4.5 for 4-bit formats, which include scaling data.
KV cache grows with context length; we assume a typical grouped-query attention layout for a model of this size. Overhead adds 0.5 GB plus 5% of the weights for runtime buffers.
Model sizes come from each model's published Hugging Face weights, matched through the OpenRouter catalog (or the publisher's model card where the weights are packed). Closed models are shown as API-only, and models without a verified size say so instead of guessing.
Unified-memory machines share memory between CPU and GPU; we assume about 75% is available to the model.