►Recent Highlights from the Previous Thread: >>105991463

--Papers:
>105993415 >105993536
--Running Qwen3-235B locally with consumer hardware and offloading tweaks:
>105992755 >105992783 >105992794 >105992802 >105992830 >105992846 >105992847 >105992860 >105992967
--Qwen3-Coder-480B-A35B-Instruct released with FP8 and 1M context claims:
>105992135 >105992190 >105992181 >105992207 >105992229
--Qwen3-235B suffers quality collapse and strange punctuation bugs at long context:
>105993502 >105993846 >105993985 >105993538 >105993585 >105993604 >105993805
--Qwen pricing and benchmark performance versus real-world utility debate:
>105992307 >105992355 >105992427 >105992539 >105992664 >105992800
--Optimizing SillyTavern with Illustrious-based NSFW models:
>105993230 >105993241 >105993293 >105993343 >105993605
--Anthropic criticized over Claude Sonnet 4 coding performance vs specialized models:
>105992892 >105992928 >105992971
--Airi Godot app updated with Viona and custom model support:
>105993116 >105994085 >105994430 >105994479 >105994911 >105994507
--Frustration over faster MLX quant availability:
>105992281 >105992289
--Speculating on the endgame of LLM development and cultural differences in AI progress:
>105991722 >105991797 >105991969 >105992815 >105992567 >105992790
--OpenAI's politicized benchmarking of Chinese LLMs draws skepticism and mockery:
>105991754 >105991859
--Model recommendations additions:
>105994067 >105994153
--Quadratic pricing for context length in Qwen3-coder models:
>105993326
--Microsoft poaches DeepMind talent:
>105991504 >105991514
--Qwen3-Coder: coding-specialized large language model:
>105991759
--Hallucination in IQ4_XS when generating lyrics without prefill:
>105991933
--Miku (free space):
>105991494 >105991507 >105991610 >105991691 >105991859 >105993126 >105993168 >105993235 >105993939 >105994056 >105995018

►Recent Highlight Posts from the Previous Thread: >>105991883

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