
3 local AI drops: 1) Patch llama.cpp to b12532+ for CVE-2026-126974 (CVSS 7.8 GGUF RCE); 2) HF Qwen3.8-27B-pi-adapted LoRA lifts SWE-bench Pro 39.2%->64.7% with thinking OFF (4.5x less tokens); 3) MiniMax-H3-Image 16MP in 12GB https://huggingface.co/unsloth/MiniMax-H3-Image-GGUF
