Fine-tuning an LLM changes how the model behaves on your examples: its tone, its output format and its style. Unsloth Studio runs it in the browser: pick a model, upload a dataset, start a LoRA or QLoRA run and export the result to GGUF for Ollama. Unsloth keeps memory use and training time low, so a model with 7–8 billion parameters trains on a single card.
In short
Unsloth Studio runs on cards with 16 GB or more. The most affordable suitable card right now is the RTX A4000 16 GB from 13 ₽ per hour, billed by the minute.
Updated 13 October 2026
A 7–8B model in 4-bit (QLoRA) fits into 16 GB, and an RTX 4090 24 GB leaves room for long examples and 14B models. Train 27–32B models in 4-bit on 48 GB, for example on an L40S 48 GB. LoRA in full precision on a 7B model needs 24 GB. Long dialogues also take memory for intermediate activations, so check the length of your examples.
Prices and availability come from the live catalog and update automatically.
JSONL is the easiest format: each line is one dialogue in the messages format. For a change of style or format, a few hundred well-chosen examples are usually enough. Quality matters more than quantity: remove duplicates and answers that are wrong or contradict the rest, or the model will learn the mistake too.
{"messages": [{"role": "user", "content": "How do I return an item?"}, {"role": "assistant", "content": "Write to support with your order number, and we will tell you what to do next."}]}Rough guide for QLoRA (weights in 4-bit) and for LoRA with weights in 16-bit:
| Model | Method | Video memory |
|---|---|---|
| 7–8B: Qwen, Llama, Gemma | QLoRA, 4-bit | 16 GB or more |
| 14B | QLoRA, 4-bit | 24 GB or more |
| 27–32B | QLoRA, 4-bit | 48 GB or more |
| 7–8B | LoRA, 16-bit | 24 GB or more |
To work out the memory for your model and context length, see the guide how much video memory an LLM needs.
GGUF is the format that Ollama and llama.cpp read. Download the exported file, write a Modelfile with one FROM line that points to it, and build the model with commands in a terminal over SSH:
# Modelfile
FROM ./my-model-q4_k_m.gguf
# in the terminal
ollama create my-model -f Modelfile
ollama run my-modelThe file name depends on your export, so use your own. If Ollama already runs on the Open WebUI + Ollama template, move the file there and run the same commands.
For a few hundred examples and one or two epochs, training a 7–8B model on an RTX 4090 usually takes from a few minutes to half an hour. More examples and epochs take proportionally longer. The result, including the GGUF export, is saved to the server's disk, and you can download it through JupyterLab.