unbelievable 🤯 the bonsai 2 build i packaged for small gpus just passed 32,000 downloads in 20 days qwen 3.8 27b on the gpu you might own, rtx 3060, 3070, 4060 ti, 4070, 5060 ti, 5070, anything rtx 30, 40 or 50, with prebuilt llama.cpp in the repo so all you need is the driver, linux or wsl2 > rtx 3060 12gb: 50.1 tok/s with my speed head at 131k context, the lean file stretches it to 196k > rtx 5060 ti 16gb: 67.3 tok/s with the full 262k window and vision loaded at the same time > the speed head doesn't change the answers, greedy output is byte identical with it on or off > 6.3gb lean file or 7gb fat file, vision with prismml's mmproj thank you to everyone who pulled it and sent their rows, owners from 8gb up to the 5070 ti already added numbers to the repo. if you haven't run it yet, tonight's the night