Key Features
- 2x faster training: Up to 70% reduced VRAM across 500+ models
- Unsloth Studio: Web UI for local model management, inference, and training
- Data recipes: Automatically create training datasets from PDFs, CSVs, DOCX files
- Model support: Qwen 3.5, DeepSeek, gpt-oss, Llama 3.1/3.2, Gemma, Mistral, Phi-4
- Reinforcement learning: GRPO with 80% less VRAM
- Multi-format: GGUF, LoRA, safetensors download and execution
- Tool calling: Self-healing tool use with code execution sandbox
- Multi-file upload: Images, audio, PDFs, DOCX for multimodal workflows
Agent Integration
AI coding agents can use Unsloth to set up local model training pipelines, configure fine-tuning jobs, and manage model inference. Agents can automate dataset preparation using data recipes and orchestrate training workflows across multiple GPUs.
FAQ
Q: What is Unsloth? A: Unsloth is an open-source tool with 58.7K+ stars for running and training AI models locally. It provides 2x faster training with 70% less VRAM, a web UI, and supports 500+ models including Qwen, DeepSeek, Llama, and Gemma.
Q: How do I install Unsloth?
A: Run curl -fsSL https://unsloth.ai/install.sh | sh on macOS/Linux, or irm https://unsloth.ai/install.ps1 | iex on Windows. For the Python library only: pip install unsloth.
Q: Which models does Unsloth support? A: Qwen 3.5, DeepSeek, gpt-oss, Llama 3.1/3.2, Gemma, Mistral, Phi-4, and 500+ other open-source models in GGUF, LoRA, and safetensors formats.