Unsloth

Unsloth란 무엇인가요?

Unsloth is an open-source framework and local web UI for training, running, and exporting open models from one unified interface. The product is centered on keeping model workflows on your own hardware, with documentation that covers local inference, fine-tuning, data preparation, and export.

The Studio release adds a no-code local workflow for running GGUF and safetensors models, preparing datasets from files such as PDFs and CSVs, and training models with built-in observability. The site positions Unsloth as a way to work with 500+ models across text, vision, audio, and embeddings without leaving a local environment.

Unsloth(으)로 무엇을 할 수 있나요?

Local model running

Run GGUF and safetensors models locally, compare two models side by side, and use uploaded images, documents, audio, or code files in the workflow.

Dataset preparation

Auto-create datasets from PDFs, CSVs, JSON, DOCX, TXT, and YAML sources, then clean and refine them in the Data Recipes workflow.

Local training controls

Train and fine-tune 500+ models with LoRA, 4-bit, 16-bit, FP8, and pre-training options, with live observability for loss, gradient norms, and GPU usage.

Tool calling and code execution

Use self-healing tool calling, web search, Bash, and Python execution, including sandboxed code runs for testing and verification.

Model export

Export models to GGUF or safetensors for use with llama.cpp, vLLM, Ollama, LM Studio, and similar runtimes.

API access

Connect to Unsloth through an API endpoint or use it with external tools such as Claude Code and Codex, as described in the Studio docs.

사용 사례

“Fine-tune models from local documents”

Prepare a local fine-tuning workflow by importing source files, converting them into datasets, and starting training with presets or custom settings.

“Local inference and evaluation”

Run a model offline on your own machine, compare outputs from two models, and test tool calling or web search in a controlled environment.

“Dataset creation and cleanup”

Turn unstructured files into training-ready datasets with Data Recipes, then refine the dataset in a visual node workflow before training.

“Model export for downstream tools”

Export a trained or base model to GGUF or safetensors so it can be used in runtimes such as llama.cpp, vLLM, Ollama, or LM Studio.

“Local API integration”

Use the API endpoint or connect external tools to a local model when you want local inference inside developer workflows.

자주 묻는 질문

What is Unsloth Studio?

Unsloth Studio is a local web UI for training, running, and exporting open models. The docs describe it as an open-source, no-code interface that works offline on your own machine.

Which operating systems does it support?

The docs say Unsloth Studio works on Windows, Linux, WSL, and macOS. It also supports NVIDIA GPUs for training, while CPU-only setups are limited to chat inference and Data Recipes.

What kinds of models can it handle?

Unsloth supports running and training models locally, exporting models to safetensors or GGUF, and using an API endpoint. The docs also mention support for text, vision, audio/TTS, and embedding models.

Does Unsloth have paid plans?

Yes. The pricing page includes a Free plan, a Pro plan, and an Enterprise plan. The free version is open source, while paid plans add higher-performance and enterprise capabilities.

Does Unsloth collect user data?

The Studio docs say Unsloth does not collect usage telemetry. It only collects minimal hardware information needed for compatibility, and Studio runs 100% offline and locally.