Overview llama.cpp 0.4.0 adds initial Qwen3.8-Flash-Next and Nemotron-3-Puzzle support, on-demand tensor reading, per-slot server context limits, video input options, and a ggml update to 0.23.0 wi...

--n-cpu-ffn sounds interesting for dense models, I might have to try that maybe with Gemma 4 12b in my pitiful 8GB VRAM lol