What is runs-on-a-3060?

runs-on-a-3060 is a collection of benchmark logs and exact commands showing what speed you can actually get running local AI language models and image models on a single RTX 3060 graphics card with 12GB of memory. It is explicitly a reference log, not a tutorial or a blog post, aimed at people who want real numbers before trying a model themselves. The repository documents one specific test machine: an RTX 3060 with about 12GB of usable memory, a Ryzen 9 9900X processor, 64GB of fast system memory, and Ubuntu Linux. For models too large to fit fully on the graphics card, the project uses a technique that offloads some of the model to system memory instead, and it consistently logs settings like cache type and thread count so results can be compared and reproduced. Results are organized into a main results table plus per-model folders with the full commands used, and separate sections for image generation models, the different software runtimes tested such as llama.cpp and several of its forks, and specific techniques like key-value cache tuning, memory offloading for large models, and power limiting. A highlights table on the front page gives quick numbers for several language models, showing both prompt processing speed and text generation speed in tokens per second, alongside one image generation benchmark. The project also includes a checklist of exactly what to record when logging a new benchmark, such as the runtime version, GPU layer settings, cache types, prompt size, generation speed, and peak memory used, so results stay consistent across entries. A short recipe page shows how to combine one of the faster techniques with a coding assistant workflow. The project is released under the CC-BY-4.0 license and accepts outside contributions.