►Recent Highlights from the Previous Thread: >>106206560

--Batch size tuning drastically improves inference speed on large models:
>106208503 >106208523 >106208554 >106208716 >106208857
--Long-context models degrade well before reaching max advertised limits:
>106210500 >106210556 >106210620 >106211141 >106210611 >106210639 >106210784
--GPT-OSS-20B hallucinates domino problem over 5000 times in 30k tokens:
>106209258
--Debian 12 upgrade breaks setup; debate over systemd, OOM, and containerized alternatives:
>106209519 >106209549 >106210096 >106209578 >106209586 >106209617 >106209630 >106209654 >106209604 >106209616 >106209644
--1M context achieved, with native support debate:
>106207420 >106207434 >106207438 >106207448 >106212202
--Running Mantella mod on low-end hardware with quantized local models:
>106208028 >106208050 >106208070 >106208078
--Seeking simplest local chatbot setup with model swap and GUI:
>106211380 >106211391 >106211412 >106211425 >106211434 >106211502 >106211509 >106212060 >106211436 >106211469 >106211479 >106211489 >106211477 >106211446 >106211493 >106211506 >106211527
--Testing base models on OpenRouter reveals parameter and endpoint limitations:
>106207105 >106207196 >106207204 >106207259 >106207263 >106207273 >106207281 >106207325 >106207345 >106207353 >106207219
--Using GLM-4.5-Air-base for explicit roleplay and troubleshooting generation issues:
>106207647 >106207657 >106207681 >106207686 >106207693 >106207723 >106207744 >106207767 >106207790 >106207802 >106207838
--Local RAG effectiveness and tooling limitations for personal knowledge retrieval:
>106206814 >106206860 >106211677
--Miku (free space):
>106207647 >106208882 >106209027 >106209040 >106209293 >106209305 >106209500 >106209636 >106209720 >106209734 >106209920 >106210313 >106210485 >106210562 >106210605 >106210679 >106210707 >106210797 >106211308

►Recent Highlight Posts from the Previous Thread: >>106207154

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