Gemma 3 27B Uncensored Runs on a Single 24GB GPU

Looking to run a large multimodal model on hardware you already own? A 27B open-weight model and its uncensored fine-tune will fit on one 24GB card.

It reads text and images at the same time, holds a 128k-token context window, and takes up roughly 18GB to 28GB of disk space. Here is what it handles and what it needs to run.

What Happened

Gemma 3 27B Uncensored is a mid-size, open-weight multimodal model. It takes text and image input simultaneously, so you can hand it a screenshot, a scan, or a photo and ask about what is in it.

The uncensored fine-tune throws out the tight filter Google ships by default. That means it will work through any visual document you put in front of it instead of stopping at a built-in refusal.

What it needs:

• VRAM: one RTX 3090 or 4090 with 24GB. It runs comfortably in FP8 or a 6-bit quant.
• System RAM: 32GB to 64GB.
• Storage: about 18GB to 28GB.
• Context window: 128k tokens.

Why It Matters

A 27B multimodal model used to mean a rented cluster or a cut-down version. This one fits on a single consumer card, which changes the math if you were planning to pay per token for image work.

The 128k context is the other half of it. Long documents, multi-page scans, and big screenshots can go in whole instead of being chopped up first. The uncensored build is what makes that useful for document analysis rather than a demo.

If you already own a 3090 or a 4090 and have the RAM to match, this is one you can actually run at home.