►Recent Highlights from the Previous Thread: >>106113484

--Paper (old): When Bad Data Leads to Good Models:
>106119129 >106119412
--MoE efficiency vs dense models under sparsity, hardware, and deployment constraints:
>106114397 >106114859 >106114920 >106115069 >106116048 >106116070 >106116124 >106116084 >106116548 >106116593
--Alleged benchmaxxing in MindLink-72B via Qwen2.5 base with test contamination concerns:
>106113679 >106113776 >106113807 >106117179 >106117203 >106117222
--XBai-o4 32B model claims and skepticism over novelty and performance:
>106116827 >106116886 >106116863 >106116920 >106116942 >106116978 >106117065 >106117106 >106117125 >106117194 >106117141 >106117142 >106117154 >106117164
--Debate over leaked model context length and training strategies for long-context LLMs:
>106117295 >106117317 >106117367 >106117621 >106117701 >106117924 >106118109 >106118182 >106118311
--Determining context size and model loading limits:
>106113641 >106113669 >106113709 >106113714 >106113765 >106113775 >106113791 >106113814 >106113839 >106113857 >106114887 >106114993 >106113689
--Future of dynamic parameter scaling in MoE architectures:
>106113836
--Debate on whether LLM plateau stems from data exhaustion or suboptimal training and filtering:
>106118310 >106118322 >106118324 >106118325 >106118329
--Phi-4's excessive reasoning loops waste tokens and frustrate users:
>106114878 >106114939 >106114995 >106116206 >106116288 >106116277
--Horizon Alpha/Beta models show strong NSFW filtering and possible red teaming via user prompts:
>106114882 >106114903 >106115173 >106115377
--New GLM-4.5 MoE pull request lands amid skepticism and hype cycles:
>106113884 >106113968 >106113992 >106114043 >106114050 >106114467 >106115095 >106115332
--Miku (free space):
>106113767 >106114066 >106114076 >106114153 >106114457 >106114483 >106117524 >106119399

►Recent Highlight Posts from the Previous Thread: >>106114309

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