How we calculate fit and speed
Every verdict on this site comes from a simple formula you can check yourself. Here it is, with its limits.
Memory fit
Weights GB = parameters (billions) × 4 × quantization memory multiplier (relative to FP32). Required = weights × (1 + 0.2) to cover KV cache at 4096-token context.
Usable memory = 90% of VRAM for discrete GPUs, 75% of unified memory for SoCs. The rest goes to the OS, the display, and the runtime.
Price per GB = listed price ÷ total memory. It is a rough value signal, not a quote.
Verdicts
- ✓ Runs: required ≤ usable memory.
- ◐ Partial offload: required ≤ 2 × usable memory.
- ✕ Does not fit: otherwise.
Measured vs estimated
Tokens/sec figures come from community benchmarks and are labeled as measured. When a device and model pair has several results, the comparison page shows the fastest one. Runtime, context length, and settings all change the result, so treat these numbers as a guide rather than a guarantee.
Fit verdicts and memory figures are estimates from the formulas above. Real usage can differ by a few GB depending on the runtime and context length.