FP8<\/strong> stores weights as 8-bit floating-point numbers<\/strong>. Compared with the original BF16 \/ FP16 (16-bit), the files and memory needs are roughly halved. Unlike 8-bit integers (INT8), it can represent a wide range of values from very small to large, so it loses little quality. On Hugging Face, FP8 models come as repositories with Source: vLLM’s quantization documentation (hardware compatibility table)<\/a> (as of 2026-09-27).<\/em><\/p>\n Local Model Watch has published 4 article(s) on models available in FP8: 1 where the repository itself is in FP8, and 3 where we found a FP8 build of the model. The lists below only include builds we have checked (the publisher’s organization and well-known quantizers); a model missing here may still have a FP8 build elsewhere. Part of our model format index<\/a>.<\/p>\nFP8<\/code> in the name (containing .safetensors<\/code> files), and publishers increasingly release official FP8 versions themselves<\/strong>.<\/p>\nWhy It Matters<\/h2>\n
\n
Tips for Running It Locally<\/h2>\n
\n
Our Coverage and Data<\/h2>\n
Main Engines That Load This Format<\/h2>\n