unsloth: Cut VRAM by 70%, Train Local LLMs 2x Faster

Unsloth is an integrated desktop application for running, training, and deploying AI models in local environments. It supports various model types, including LLMs, diffusion, embeddings, and audio models, allowing direct interaction with the latest models such as Kimi K3, Qwen3.8, DeepSeek-V4, and Gemma 4. It offers broad hardware compatibility, supporting CPU, Apple, NVIDIA, AMD, Intel, and multi-GPU environments.

Model training is 2x faster than traditional methods, with VRAM usage reduced by up to 70%. This process supports various training techniques without accuracy loss, including LoRA, QLoRA, full fine-tuning, reinforcement learning, GRPO, DPO, and FP8. Trained models can be exported or deployed in GGUF, NVFP4, and FP8 formats, making them easy to use in diverse environments.

It provides features to connect external agents and tools like Claude Code, Codex, and MCP to local models. A single 'unsloth start' command can link agents to local models, and model services can be exposed externally via an OpenAI-compatible API. Additionally, the Data Recipes feature generates datasets from PDF, CSV, and DOCX files, enabling RAG and web search tasks.

It is available in three forms: a Tauri-based native desktop app, a web-based UI called Studio, and a code-based Core. Notably, Studio runs on Windows, Linux, and macOS, and supports secure HTTPS remote access via Cloudflare tunnels. It is ideal for developers who want to directly control AI models locally and build training pipelines.