oguzeldereli/ttf-ternary: Train LLMs with less memory on consumer GPUs

An AI researcher has developed a new optimizer that significantly reduces the VRAM required to train large language models, dropping memory usage from around 16 bytes per weight down to roughly 1 byte. The author fitted the training of a 27B parameter model onto a single 24GB consumer GPU and successfully converted and trained existing models like Qwen 3.8. Code, papers, and quantized GGUF weights have been released open-source for non-commercial use. Open article →