Unsloth Studio on a rented GPU
Fine-tune open models on a cloud GPU
Pick a GPU and it starts with Unsloth Studio installed: train Llama, Qwen or Gemma on your own data from the browser, chat with the result and export it to GGUF. A password for the rental comes with the link.
- H200 NVL 140 GBfrom $3.50/h
- RTX A6000 48 GBfrom $0.40/h
- RTX 5090 32 GBfrom $0.63/h
- RTX 4090 24 GBfrom $0.45/h
- RTX 4070 Ti SUPER 16 GBfrom $0.25/h
Measured, not quoted
On a rented RTX 4090
Measured on an RTX 4090 rented on ParalonCloud with Unsloth's standard QLoRA recipe on the Alpaca set. Times scale with the length of your examples.
What you get
Installed, running, yours
Unsloth Studio, latest release
Installed by Unsloth's own installer at its latest version, with its own Python and PyTorch.
Training, chat and export
Fine-tune with LoRA or QLoRA, try the model in the same app, export to GGUF for Ollama or llama.cpp.
Up to 70% less VRAM
An 8B model fine-tunes in 6.3 GB, so cards well below a 4090's size do the job.
Its own password, per rental
Unsloth Studio always asks for a login: the username and this rental's password come with its link.
Which GPU
What fits on which card
| Model (QLoRA) | 12 GB | 16 GB | 24 GB (RTX 4090) | 32 GB (RTX 5090) |
|---|---|---|---|---|
| 1B to 4B | Comfortable | Comfortable | Comfortable | Comfortable |
| 7B to 8B (Llama 3.1 8B, Qwen3 8B) | Fits (6.3 GB measured) | Comfortable | Comfortable | Comfortable |
| 12B to 14B | Tight | Fits | Comfortable | Comfortable |
| 27B to 32B | No | No | Fits | Comfortable |
Scaled from the 6.3 GB measurement; longer examples need more. More in Fine-tune an LLM with Unsloth Studio.
How it works
Three clicks to Unsloth Studio
- 01
Pick a GPU for your model
The list filtered for Unsloth Studio shows machines with 12 GB of VRAM or more and 60 GB of disk.
- 02
Rent with Unsloth Studio
The machine answers in about a minute; Studio installs on top, a few minutes since it downloads 4.5 GB.
- 03
Sign in and train
Copy the username and password shown with the link, pick a base model and your dataset, and start.
FAQ
Unsloth Studio in the cloud
How much does it cost to fine-tune an LLM with Unsloth?
The GPU by the minute. On an RTX 4090, Llama 3.1 8B with QLoRA trained 2,000 examples in about 4 minutes, about 3 cents; the full Alpaca set of 51,760 in about 1.8 hours, about 81 cents.
Which GPU do I need for Unsloth?
12 GB of VRAM or more. QLoRA fine-tunes an 8B model in 6.3 GB; 27B to 32B models need a 24 GB card or bigger.
What is the password?
Unsloth Studio always has a login. The username is unsloth and the password is made for each rental; both are shown next to the Open button, and nowhere else.
How do I get my fine-tuned model out?
Export it from Studio to GGUF or merged weights and download it, or push it to Hugging Face, before you stop: everything on the machine is deleted when the rental ends.
Which Unsloth version do I get?
Unsloth's latest release, installed by its own installer when the machine starts.
Can I start it from code or an AI agent?
Yes: "template": "unsloth" on POST /api/v1/rentals, or ask an agent with the ParalonCloud MCP server. The rental gives Studio's link and its sign-in.
Unsloth Studio, ready in minutes
From $0.25 an hour, billed per minute. Stop whenever you like.