Data & method

• Model sizes (real): parameter counts from the Hugging Face API (safetensors metadata), retrieved 2026-10-17.
• Measured speeds (real): community llama-bench results (tg128 = generation, pp512 = prompt processing, Llama 2 7B) parsed from llama.cpp scoreboards: Apple Silicon #4167, CUDA #15013, Vulkan #10879. 763 rows in total.
• Newer models (real): llama-bench tables posted in llama.cpp discussions (gpt-oss guide #15396, #17621 and others) where the device is stated, plus public LocalScore results (Llama 3.1 8B and Qwen2.5 14B, Q4_K_M, llamafile). Every row links to the exact post or result. We never invent benchmark numbers.
• Contribute a benchmark: submit your llama-bench output here (reviewed before it appears), or post your llama-bench output (with the GPU name) in the llama.cpp discussions or run LocalScore; we pull both on refresh.
• Multi-GPU & offload (estimate): VRAM is summed across GPUs for the fit verdict; speed is not additive. "Runs slowly with offload" means the model fits in VRAM plus system RAM minus ~6 GB for the OS.
• Hardware memory (spec): manufacturer VRAM/unified-memory sizes.
• Memory needed (estimate): params × bits-per-weight (Q4_0 4.5, Q4_K_M 4.85, Q5_K_M 5.69, Q6_K 6.56, Q8_0 8.5, F16 16) ÷ 8, plus 10% and 1 GB for a ~4k context.
• Speed for other models (estimate): measured Llama 2 7B Q4_0 speed on the same chip, scaled by weight size. Not done for MoE models.