# Run Llama 3.1 8B

The usual starting point for teaching a model your own writing or data. Ready in about 90 seconds on an RTX 4090, from $0.45/hr. No install, no code.

## What it does

Llama 3.1 8B is small enough to train on a single card and capable enough to be useful afterwards. That combination is why most fine-tuning starts here.

Fine-tuning means showing the model a few hundred to a few thousand examples of the work you want, so it learns your tone, your format, and your vocabulary. It does not teach the model new facts reliably. It teaches it a job.

Typical results: a support reply writer that sounds like your team, a summarizer that produces your report format, a classifier that uses your internal category names.

## Your setup

We run Llama 3.1 8B on an RTX 4090 with 24GB of memory, at $0.45 per hour, billed by the minute. It takes about 2 seconds per example.

Training needs far more memory than running, because the card has to hold the model, a copy of the gradients, and the optimizer state at the same time. Full fine-tuning of an 8 billion parameter model needs around 80GB. We use LoRA instead, which trains a small adapter rather than the whole model and brings the requirement down to about 16GB. That fits an RTX 4090 with room to spare, and for the kind of style-and-format learning most people want, LoRA gets you the same result.

## What it costs

1,000 training examples takes about 40 minutes of GPU time, which is about $0.30.

## Two ways to run it

- **Open the app.** JupyterLab with Axolotl installed and a working config already filled in. For people who want to see the training curve.
- **Run a batch job.** Upload a spreadsheet of examples, set the strength, and collect a finished model. This is the path most people should take.

## Model facts

| Field | Value |
|---|---|
| Parameters | 8B |
| License | Llama 3.1 Community License |
| Memory required | 16GB minimum for LoRA, 80GB for full fine-tuning |
| Base model | Trained from scratch by Meta |
| Publisher | Meta |
| Recommended hardware | RTX 4090, 24GB |
| Category | [Fine-tuning](https://gpuvault.io/working-on/fine-tuning/) |

## Questions people ask

### How many examples do I need to fine-tune a model?

Between 200 and 1,000 for most tasks. Below 200 the model tends to memorize rather than learn a pattern. Above about 2,000 the returns flatten out for style-and-format work. Quality matters far more than quantity: 300 carefully checked examples beat 3,000 messy ones every time.

### Who owns the model I train?

You do. The adapter you produce is yours, it downloads to your machine, and we do not use it or your training data for anything. The base model keeps its original license, so a model fine-tuned from Llama 3.1 is still subject to the Llama Community License when you deploy it.

### Do I need to install anything to use Llama 3.1 8B?

No. We start a machine with Llama 3.1 8B already loaded and hand you a link. Everything runs in your browser, there is nothing to download, and nothing is left on your computer afterwards. It works the same on a Mac, a Windows laptop, or a Chromebook. A workspace is usually ready in about 90 seconds.

### How much does it cost to run 1,000 training examples?

About $0.30. Llama 3.1 8B takes about 2 seconds per example on the RTX 4090 we recommend, so 1,000 training examples is roughly 40 minutes of GPU time at $0.45 per hour. Billing is by the minute, and a batch job shuts the GPU off the moment the last item finishes. The calculator above works this out for your own numbers.

### Can I run Llama 3.1 8B on a cheaper card?

Sometimes, and the calculator will not always make it look worth it. Training needs far more memory than running, because the card has to hold the model, a copy of the gradients, and the optimizer state at the same time. Full fine-tuning of an 8 billion parameter model needs around 80GB. We use LoRA instead, which trains a small adapter rather than the whole model and brings the requirement down to about 16GB. That fits an RTX 4090 with room to spare, and for the kind of style-and-format learning most people want, LoRA gets you the same result. If you want to try a different card anyway, the advanced catalog lets you pick one and shows the estimated time before you commit.

### What are the license restrictions on Llama 3.1 8B?

Llama 3.1 8B is released under the Llama 3.1 Community License. That is not a permissive license, so read it before you use the output in paid work. We show the license on every model page precisely because this catches people out, and there is usually a permissively licensed alternative in the same category.

---

*Source: https://gpuvault.io/model/llama-3-1-8b/ — GPUVault, GPU rental by the minute.*