Pick a model and your hardware. We compute weights for every GGUF quant (Q2_K → Q8), KV cache at your context length, and a tokens/sec estimate from memory bandwidth. Parameter counts pulled live from HuggingFace safetensors metadata.