►Recent Highlights from the Previous Thread: >>105863705

--Paper: Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation Is Wasteful:
>105864019 >105864154 >105867290
--Kimi K2 MoE model release sparks debates on local hosting, performance, and the future of large language model scaling:
>105870772 >105870794 >105870780 >105870785 >105870789 >105870790 >105870832 >105870849 >105870851 >105870875 >105870838 >105870837 >105870847 >105870879 >105870912 >105870915 >105870926 >105871087 >105871584 >105871630 >105871643 >105870946 >105870958 >105870964 >105870973 >105870987 >105871813 >105871815
--DeepSeek-R1-0528 system prompt support and rendering behavior clarified:
>105864170 >105864191 >105864222 >105864339 >105864436 >105864457 >105864469 >105864507 >105864814
--Accusation of Nvidia deliberately restricting GPU performance in drivers unless functions use "cutlass_" prefix:
>105869938
--Tradeoffs in scaling large MoE models and impact of safety restrictions on release timelines:
>105863885 >105864003 >105864059 >105864102 >105864248 >105864286 >105864465 >105864483 >105864523 >105864564 >105864106 >105864175 >105864233
--Grok4 reception and technical challenges of running large models locally with limited resources:
>105864963 >105865011 >105865051 >105869354 >105865410 >105865527 >105865544 >105865638 >105865923
--Jamba mini underperforms in roleplay and long-context comprehension despite low censorship:
>105870365 >105870410 >105870623 >105870699
--Status update on pending llama.cpp row parallelism feature implementation:
>105870286 >105870423
--Granite 4 (Mamba2 MoE) support merged into llama.cpp:
>105867175
--Logs: Kimi-K2:
>105871284 >105871342 >105871480 >>105871729 >105871652 >105871755 >105871773
--Miku (free space):
>105864655 >105868025 >105869430

►Recent Highlight Posts from the Previous Thread: >>105863712

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