Benchmarks & research
TurinTech's Artemis files compress Clef, Decider and Kev for llama.cpp, balancing memory savings against agreement with the original models' answers.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
These are independent compressed derivatives, not Jev weights or endorsements by the original makers. TurinTech reports how often they keep the source model's answer, which can also preserve its mistakes. Because the evaluation helped select the files, it does not establish how they perform on unseen inputs.
Pick a file that fits your memory and use its documented decision service, rather than a chat-model button. Clef needs both batch settings at 4,096 or higher. A lower setting can reject a long request even when the model's own input limit would allow it.