►Recent Highlights from the Previous Thread: >>105872817

--Speculative future AI architectures and the limits of self-modifying models:
>105874825 >105874937 >105875601 >105875782 >105876056
--Kimi shows strong roleplay performance with potential AO3-trained quality:
>105876506 >105876543 >105876600
--Kimi-K2 GGUF model released with ktransformers support and Q4_K_M quantization:
>105877806 >105877819 >105877832 >105877855
--Distinguishing censorship origins in base and instruct-tuned \models:
>105876179 >105876186 >105876207 >105876230 >105876346 >105876470 >105876491 >105876540 >105876549 >105876661
--Kimi model criticized for excessive refusals and censorship:
>105876194 >105876213 >105876237 >105876428 >105876465 >105876558
--Japanese language roleplaying advantages and model performance limitations:
>105877325 >105877332 >105877370 >105877388 >105878979 >105877352 >105878897 >105878931
--Debate over ablation's impact on model refusal behavior and alternative expert-targeted fine-tuning approaches:
>105877689 >105877703 >105877715 >105877733 >105877757 >105877762 >105877755 >105877764
--Jailbreaking techniques to bypass model restrictions on explicit content generation:
>105874973 >105875018 >105875049 >105875118 >105875361 >105875087 >105875104 >105875121
--FP8 performance gains tied to Triton kernel naming tricks:
>105873562 >105873634
--Mockery of OpenAI's delayed open-weight model and safety justification:
>105876448 >105876531 >105876561 >105876605 >105876646 >105876629
--Speculation on Meta's model experiments and critique of AI industry's environmental priorities:
>105874049 >105874083 >105874158 >105874235 >105874155 >105874191
--Voice cloning with Openaudio S1 Mini and Resemble Enhance audio cleanup:
>105877122
--Miku (free space):
>105875688 >105875887 >105876796 >105878089

►Recent Highlight Posts from the Previous Thread: >>105872822

Why?: 9 reply limit >>102478518
Fix: https://rentry.org/lmg-recap-script

>>105815838
Lubuntu is lightweight Ubuntu whereas kubuntu/xubuntu are desktop flavors.
Manjaro is a fleshed-out operating system with its own tooling (e.g. pamac) whereas Endeavour is just an Arch installer.

>>105809980
Good luck anon!
It's too late for me since I'm 28, don't make the same mistake

Post medium size (70b to 123b) model recommendations for ramlets like me who have no hope of running larger models like deepseek R1
In return I'll start with my list of models I have been using for the past few months, including smaller ones
>EVA LLaMA 3.33 v0.0 70b Q5_K_M (best)
>Luminum v0.1 123b Q3_K_M (second best)
>LLaMA 3.3 70b Instruct Q5_K_M (original model, no finetune)
>Skyfall 36b Q6_K (smart for size)
Honorable mentions:
>Mistral Nemo Instruct 12b Q8_0 (smart for size and uncensored when jailbroken, but too small)
>Gemma 3 27b Q6_K (great writing style, easy to jailbreak, but it's just too averse to vulgar language and schizos if you push it too hard)
>Cydonia 24b Q6_K (too retarded)
>Cogito v1 36b IQ4_NL (too schizophrenic)
>L3.3 TRP BASE 80 70b Q6_K (too schizophrenic)