►Recent Highlights from the Previous Thread: >>106011911

--Papers:
>106015367 >106015437 >106018967
--Magistral model requires --special flag in llama.cpp to expose [THINK] tokens for proper frontend parsing:
>106012674 >106012735 >106012780 >106012820 >106012845 >106014062 >106014161 >106012821 >106012879 >106012906 >106012953 >106012925 >106013180 >106013214 >106013298 >106013344 >106013388 >106013426 >106013500 >106013544 >106013579 >106013665
--Qwen3's over-alignment response to a Japanese slur term sparks criticism of modern LLM behavior:
>106018450 >106018461 >106018492 >106018565 >106018581 >106018621 >106018646 >106018669 >106019891 >106019901 >106019919 >106019951 >106018716 >106019655
--Mistral's flawed two-server approach breaks llama.cpp's lightweight design:
>106015554 >106015638 >106015666 >106015868 >106020302 >106020383
--Local execution challenges with large MoE models under VRAM and offloading constraints:
>106014862 >106014916 >106015424 >106015454 >106015519 >106015565 >106016265 >106017240 >106015507 >106015028
--High-thread NVMe LLM inference benchmarks reveal diminishing returns:
>106012162 >106012174 >106013567 >106013660 >106013723 >106013786 >106018346 >106013796 >106013797
--MLX outperforms llama.cpp on M3 Ultra but has SillyTavern integration issues:
>106013458 >106014077 >106014268
--Quantized Qwen3 coder model benchmarks:
>106017215 >106017260 >106017277 >106019453 >106019463 >106019547 >106020316 >106018035
--Performance regression in llama.cpp multi-GPU inference after update:
>106012278 >106013195 >106013325
--Anon extracts Higgs Audio v2 patches due to missing vLLM fork:
>106013788 >106014022 >106019174
--Misc:
>106016151 >106019426 >106020161 >106020202 >106021271 >106021313 >106018793 >106021835
--Miku (free space):
>106011969 >106012287 >106013068 >106016203 >106018731 >106018817 >106019327

►Recent Highlight Posts from the Previous Thread: >>106011918

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