LLMTools is an open-source library from Cornell University’s Kuleshov group for running and fine-tuning large language models on a single consumer GPU — even in 2-bit precision. It is built around the group’s ModuLoRA research, which makes it practical to adapt capable open models on the kind of graphics card found in a home workstation. The project is hosted on GitHub, where it has earned roughly 730 stars.
Fine-tuning LLMs normally demands datacenter accelerators because weights, gradients and optimizer states overwhelm consumer VRAM. LLMTools attacks the problem with aggressive quantization: it fine-tunes models in 2-bit, 3-bit and 4-bit precision using ModuLoRA, which integrates low-rank adapters directly with the quantized backbone so training stays memory-lean without collapsing quality. It exposes a friendly Python API covering quantization, inference and fine-tuning, with modular support for multiple models, quantizers and optimization algorithms. Fine-tuned results can be shared straight to the Hugging Face Hub. The library is a research artifact — the code behind published papers — kept usable for practitioners. It is a good illustration of how quickly quantization research moved from papers into tools that a single workstation can run.
LLMTools is free and open source. There is no commercial version or hosted service — the whole point is running serious fine-tuning on hardware you already own.
LLMTools is for researchers, students and hobbyists who want to fine-tune real LLMs but only have a consumer GPU, and for anyone studying how far quantization can push local model adaptation.