Unsloth – Train LLMs 2x faster with 70% less VRAM (github.com)

Unsloth has announced a significant update that enables the training of large language models (LLMs) at twice the speed with drastically reduced VRAM requirements—up to 80% less for specific models. This breakthrough is crucial for the AI/ML community as it allows developers to train sophisticated models like GPT and LLaMA on consumer hardware, making advanced AI more accessible to researchers and hobbyists alike. For instance, models like gpt-oss and Llama can now function efficiently on just 14GB VRAM while delivering impressive performance metrics, such as faster training speeds compared to previous methods. The update introduces features like Quantization-Aware Training and innovative memory-efficient reinforcement learning algorithms, which optimize both speed and accuracy without approximating methods. Additionally, support for a broad array of models—including vision and text-to-speech—along with contextual fine-tuning for over 500K tokens, presents new possibilities for developers and researchers in enhancing language models. The implementation of a Docker container and streamlined installation process further reduces barriers to entry, empowering more users to leverage AI technologies effectively.