overview
What is unsloth?
unsloth is an open-source framework developed by Unsloth that enables developers, ML practitioners, and beginners to fine-tune and run large language models locally. It significantly accelerates and optimizes LLM fine-tuning, making the process faster and more memory-efficient without compromising accuracy. The platform, including Unsloth Studio, is a Python library and web UI designed to streamline the fine-tuning of LLMs such as Llama (versions 1, 2, 3), Mistral, Gemma, and Phi-3. This is achieved through advanced mathematical derivations, custom GPU kernels written in OpenAI's Triton language, and optimized techniques like Flash Attention 2, LoRA, and QLoRA.
Recent developments, as of March-April 2026, include Unsloth Studio enhancements such as seamless Windows CPU/GPU support, pre-compiled llama.cpp binaries for 6x faster installs and 50% reduced installation sizes, and improved tool calling. Gemma 4 support was updated in April 2026, with Unsloth training Gemma 4 approximately 1.5x faster with about 60% less VRAM than traditional setups. Further updates in October 2025 introduced faster and more memory-efficient Reinforcement Learning (RL) with 50% less VRAM and 10x more context, alongside Unsloth Flex Attention for gpt-oss enabling over 8x longer context and over 50% less VRAM.