Unsloth
Unsloth Studio is a local web UI for running and fine-tuning language models. It can search, download and serve GGUF, LoRA and safetensors models, compare two models side by side, export to other formats, and expose OpenAI- and Anthropic-compatible endpoints so local models can be used from Claude Code, Codex and similar tools. Fine-tuning with LoRA is configured entirely through the UI.
Unsloth
In Development
This script is currently in active development and may be unstable or incomplete. Use in production environments is not recommended.
This script is currently in active development and may be unstable or incomplete. Use in production environments is not recommended.
Installation
Default install:
1
bash -c "$(curl -fsSL https://raw.githubusercontent.com/community-scripts/ProxmoxVED/main/ct/unsloth.sh)"
Notes
Log in as the user 'unsloth'; the generated password is in ~/unsloth.creds inside the container. Rotate it with 'unsloth studio reset-password', which prints a new one and needs no restart.
Web search only works with an external provider (OpenAI, Anthropic, OpenRouter, Gemini or Kimi) configured in the settings, because the search runs server-side at the provider. Local GGUF models do not offer it and the search button stays hidden for them.
Fine-tuning requires an NVIDIA GPU. Without one, inference, model management, export and the OpenAI-compatible API work, but training does not. Pass a card through and install the NVIDIA driver in the container to enable it.
Models and Hugging Face caches land in /opt/unsloth_data, which fills up quickly - a single 7B model at Q4 is roughly 4 GB. Give the container more disk if you plan to keep several, and back that directory up rather than /opt/unsloth.
PyTorch is installed with uv's --torch-backend=auto, so a CPU-only container gets the CPU wheels and a GPU container the matching CUDA build. Re-run the update after adding a GPU so the right backend is pulled.