The multilingual e5 family is one of the best options for multilingual embedding models. This is the GGUF version of https://huggingface.co/intfloat/multilingual e5 large instruct. Check out their prompt recommendations for different tasks! It is supported since the XLMRoberta addition in llama.cpp was merged on 6th August 2024. https://github.com/ggerganov/llama.cpp/pull/8658 Currently q4 k m, q6 k, q8 0 and f16 versions are available. I would recommend q6 k or q8 0. In general you barely have any performance loss going to 8 bit quantization from base models, while there usually is a small but noticable dropoff occuring somewhere between q6 q4. At some point the dropoff gets pretty massive going towards ~q3 or lower.
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy