METHOD
Derivatives and 30-day downloads from mkhr-data hf-hub-models, built daily from public Hub metadata. "Uncensored" = repository name matches uncensored / abliterated / heretic / josiefied / decensor / jailbreak / obliterat / unalign; it can miss or over-include. Sizes are Hub file sizes (GiB); shards summed; mmproj, draft and MTP side files excluded. KV cache: f16, full-attention layers only, from the base model's config; llama.cpp allocates about 15% more. 1M needs YaRN and recall degrades long before it. i-quants are slow on Apple Metal. The 8 GB reserve is an estimate; nothing here is a measurement of your machine.
DATA
Snapshot
, Asia/Tokyo. Aggregate counts and repository identifiers only; no model weights, cards or user data are redistributed. Source dataset manifest:
hf-hub-models. Hub metadata is subject to the
Hugging Face Terms of Service; each model's weights carry its own license shown on its Hub page.
LEGAL
Published by 株式会社幕張開発 (Makuhari Development). Not affiliated with, endorsed by, or sponsored by Alibaba Cloud / Qwen, Hugging Face, Apple, or any listed uploader; names and marks belong to their owners. Listing a repository is not a recommendation; "uncensored" builds have had safety behaviour removed by third parties and are used at your own risk under their licenses. Figures are provided as-is without warranty. No cookies, no tracking, no accounts.