Tell us what machine you have — or let your browser detect what it can — and see instantly which open-weight models run on it, which are borderline, and which are out of reach. Every verdict follows the published sizing rule at the bottom of the page. Nothing is uploaded: the check runs entirely in your browser.
The sizing rule, published: memory need = weights + KV cache + runtime. Weights: a 4-bit model (Q4_K_M) weighs about 0.58 GB per billion parameters. KV cache: grows with the context you use — we estimate 0.125 MB per token at 8B, scaling with parameters at the 0.45 power (calibrated on common GQA architectures like Llama: ~1 GB at 8K for an 8B, ~10 GB at 32K for a 70B; individual architectures vary by ±40%). Runtime overhead: 0.8 GB. A model fits when the total stays under 85% of your GPU memory; 85–110% is borderline; beyond, CPU RAM may still work, slower. Apple Silicon shares one pool: we count 70% of unified RAM — and we point you to MLX builds, Apple’s optimized format. Mixture-of-experts models are sized by their total parameters. Parameter counts are read from model names; models without a readable size are excluded. The list merges our hosted-API catalog with weights-only releases from the public Hugging Face catalog (declared licence, not gated) — so local-first models with no API appear here too. Deterministic, no account, no telemetry — corrections welcome at [email protected].