To fine-tune a 7B model with Unsloth you need at least 5 GB of VRAM with QLoRA or 19 GB with LoRA. A 24 GB card covers QLoRA up to 27B and LoRA up to 9B. The full table below is Unsloth’s published minimum, checked 2026-10-22 on the Unsloth requirements page; it is not a measurement by vramcalculator.io.
Unsloth’s own caveat: “QLoRA uses 4-bit, LoRA uses 16-bit. Keep in mind that sometimes more VRAM is required depending on the model so these numbers are the absolute minimum.”
Minimum VRAM by model size
| Model parameters | QLoRA (4-bit) VRAM | LoRA (16-bit) VRAM |
|---|---|---|
| 3B | 3.5 GB | 8 GB |
| 7B | 5 GB | 19 GB |
| 8B | 6 GB | 22 GB |
| 9B | 6.5 GB | 24 GB |
| 11B | 7.5 GB | 29 GB |
| 14B | 8.5 GB | 33 GB |
| 27B | 22GB | 64GB |
| 32B | 26 GB | 76 GB |
| 40B | 30GB | 96GB |
| 70B | 41 GB | 164 GB |
| 81B | 48GB | 192GB |
| 90B | 53GB | 212GB |
| 405B | 237 GB | 950 GB |
Figures are copied exactly as printed, including the inconsistent spacing (22GB vs 3.5 GB). Nothing was rounded. A copy of the fetched page is kept at docs/evidence/unsloth-requirements-20261022.md.
What fits your card
Reading the table against common card sizes (a minimum is met when the card’s memory is at least the listed figure):
- 8 GB: QLoRA up to 11B; LoRA on 3B.
- 12 or 16 GB: QLoRA up to 14B; LoRA on 3B.
- 24 GB: QLoRA up to 27B; LoRA up to 9B.
- 32 GB: QLoRA up to 40B; LoRA up to 11B.
- 80 GB: QLoRA up to 90B; LoRA up to 32B.
Because these are minimums, a model at the edge of your card may still run out of memory. If it does, Unsloth’s advice is: “A common issue when you OOM or run out of memory is because you set your batch size too high. Set it to 1, 2, or 3 to use less VRAM.”
The GPU-by-GPU version of this list, using the exact memory of each card in the catalog, is in which GPUs meet the Unsloth minimums. For how fine-tuning memory compares with just running a model, see inference vs QLoRA VRAM.
FAQ
How much VRAM do I need to fine-tune a 7B model?
Unsloth lists 5 GB for QLoRA and 19 GB for LoRA on a 7B model. Those are minimums; some models need more.
What is the difference between the QLoRA and LoRA columns?
QLoRA loads the base model in 4-bit, LoRA in 16-bit. That is why the LoRA column is two to four times larger for the same model.
I'm running out of memory during fine-tuning. What should I change first?
The batch size. Unsloth says a too-high batch size is the common cause and suggests setting it to 1, 2, or 3.
Did vramcalculator.io measure these numbers?
No. The table is copied from Unsloth's requirements page as checked on 2026-10-22. This site has not run these fine-tunes.
Calculator version not recorded for this article. Sources last checked 2026-10-22.