FreeLLMThe best model your hardware can run
Pick your graphics card, how many you have, and your system RAM. See which top-voted open models fit, how the memory splits, and how fast they will write.
How we estimate
Memory = weights (parameters × bytes per weight at the chosen precision) + context cache (computed from each model's real architecture on Hugging Face: layers, attention heads, sliding windows) + about 0.6 GB per card for the runtime. Image and video models get 3 GB more for text encoders and decoders.
Speed: writing a token reads the active weights and the cache once, so tokens per second ≈ memory bandwidth × 78% ÷ data read, plus a fixed cost for every token (about 1 ms; 2.8 ms for mixture-of-experts models, which also pick their experts; more on Apple). Mixture-of-experts models read only their active experts. With several cards (llama.cpp and Ollama split layers across them), speed stays near one card's speed; with weights in system RAM, the RAM part is far slower. The formula is calibrated on 44 real llama.cpp tests (list with sources): typical error about ±10%, a few cases up to ±40%.
Apple, AMD Ryzen AI Max and DGX Spark share memory with the system: we count 75% of it.