Kimi K2.5 (GGUF imatrix quants)
This repo packages moonshotai/Kimi-K2.5 into a large set of GGUF files, quantized with llama.cpp using the imatrix workflow. Practically, it’s meant to make Kimi K2.5 usable in the ecosystem of “GGUF-first” runtimes (llama.cpp, LM Studio, KoboldCpp, text-generation-webui, etc.) without needing to do your own conversions. The author provides many quant variants, ranging from higher-quality but heavy options (Q6/Q8) down to more compact ones (Q4, IQ*), plus notes about online repacking behavior on ARM/AVX CPUs.
What to try first: pick a mid-range quant (often a Q4_K_M/Q5-style file if you have the VRAM/RAM) and download only that specific file or folder with huggingface-cli --include instead of cloning the whole repo. If you’re not sure what your runner supports yet, check the original model page and your tool’s “supported models” list first—new architectures often require a recent llama.cpp build.
Quick stats from the listing feed: pipeline: text-generation · 6 likes · 7453 downloads.
Source listing: https://huggingface.co/models?sort=modified