# Run Gemma 3 27B

A mid-sized assistant that reads images and handles very long documents. Ready in about 90 seconds on an L40S, from $0.89/hr. No install, no code.

## What it does

Gemma 3 27B sits between the small fast models and the large expensive ones. It is noticeably better than a 24 billion parameter model at longer reasoning, and noticeably cheaper than a 70 billion one.

It reads images alongside text and holds a context of about 128,000 tokens, which is a long report or a stack of contracts in one go.

It is a good default when you are not sure yet how hard your task is.

## Your setup

We run Gemma 3 27B on an L40S with 48GB of memory, at $0.89 per hour, billed by the minute. It takes about 29 seconds per thousand words.

Gemma 3 27B is about 16GB in its 4-bit form, which would fit a 24GB card easily. The reason we use a 48GB card is the context window. Filling 128,000 tokens of context costs roughly 20GB on top of the model, and running out of context memory mid-document is the most common way this model fails. The L40S has the headroom to avoid it.

## What it costs

150 thousand words written takes about 73 minutes of GPU time, which is about $1.08.

## Two ways to run it

- **Open the app.** A chat window with document and image upload.
- **Run a batch job.** Spreadsheet in, spreadsheet out.

## Model facts

| Field | Value |
|---|---|
| Parameters | 27B |
| License | Gemma Terms of Use |
| Memory required | 24GB minimum (4-bit), 48GB recommended for long documents |
| Base model | Trained from scratch by Google DeepMind |
| Publisher | Google DeepMind |
| Recommended hardware | L40S, 48GB |
| Category | [Language models](https://gpuvault.io/working-on/language-models/) |

## Questions people ask

### Do I need to install anything to use Gemma 3 27B?

No. We start a machine with Gemma 3 27B 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 150 thousand words written?

About $1.08. Gemma 3 27B takes about 29 seconds per thousand words on the L40S we recommend, so 150 thousand words written is roughly 73 minutes of GPU time at $0.89 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 Gemma 3 27B on a cheaper card?

Sometimes, and the calculator will not always make it look worth it. Gemma 3 27B is about 16GB in its 4-bit form, which would fit a 24GB card easily. The reason we use a 48GB card is the context window. Filling 128,000 tokens of context costs roughly 20GB on top of the model, and running out of context memory mid-document is the most common way this model fails. The L40S has the headroom to avoid it. 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 Gemma 3 27B?

Gemma 3 27B is released under the Gemma Terms of Use. 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.

### What happens to my files after I stop?

Anything you save into your workspace folder stays in your account and is there when you come back. Everything else is destroyed: the machine is terminated and its disk is wiped. Batch job inputs and outputs are deleted from our storage 24 hours after you download them. We do not read your files or use them to train anything.

### What if it fails partway through?

You are not charged for work that did not complete, and failed sessions are refunded automatically. Batch jobs write each output as it finishes rather than at the end, so a failure at item 380 of 400 leaves you 379 usable files and a message saying what broke. Restarting picks up where it stopped rather than redoing the work.

---

*Source: https://gpuvault.io/model/gemma-3-27b/ — GPUVault, GPU rental by the minute.*