Running Qwen2.5-7B-Instruct at Q4_K_M takes 4.92 to 5.42 GiB at 8K context. Unsloth lists 5 GB as the minimum for a 7B QLoRA fine-tune and 19 GB for 16-bit LoRA. So if a 7B runs on your card at Q4, a QLoRA run of the same size is in the same range, while full 16-bit LoRA needs a much bigger card.
Read the two columns as numbers from two different sources, not as one measurement. Unsloth’s figures are its own published minimums (4-bit for QLoRA, 16-bit for LoRA), which it says can rise with batch size. Ours are inference estimates built from a specific file, an 8K-token KV cache and runtime overhead. Unsloth prints GB; we work in GiB. Neither column is converted into the other.
Side by side
| Model you’d run | Inference at 8K context | Unsloth size class | Unsloth QLoRA min | Unsloth LoRA min |
|---|---|---|---|---|
| Qwen2.5-7B Q4_K_M | 4.92 to 5.42 GiB | 7B | 5 GB | 19 GB |
| Qwen3-14B Q4_K_M | 9.76 to 10.39 GiB | 14B | 8.5 GB | 33 GB |
| Qwen3-32B Q4_K_M | 20.53 to 21.41 GiB | 32B | 26 GB | 76 GB |
| Llama 3.1 8B BF16 | 16.08 to 16.83 GiB | 8B | 6 GB | 22 GB |
| Llama 3.3 70B BF16 | 134.04 to 137.22 GiB | 70B | 41 GB | 164 GB |
Inference: one sequence, FP16 KV cache, llama.cpp. Qwen2.5-7B file: 4,683,073,632 bytes. Unsloth figures copied on 22 October 2026 and stored in docs/evidence/unsloth-requirements-20261022.md.
What the table tells you
- QLoRA minimums track Q4 inference at small sizes. At 7B both sit around 5. At 14B, our inference estimate (with an 8K cache) is a bit above Unsloth’s 8.5 GB QLoRA minimum. At 32B, Unsloth’s 26 GB is above our 20.53 to 21.41 GiB. Don’t read either gap as a correction of the other.
- 16-bit LoRA is a different league. It’s roughly three to four times the QLoRA minimum at every size. Only the 7B and 8B LoRA minimums (19 and 22 GB) are below a 24 GB card’s label.
- The Llama rows aren’t a like-for-like comparison. We don’t list a Llama GGUF, so the inference column for Llama is full BF16. A 4-bit Llama would be much smaller to run. See Llama 3 VRAM requirements.
For which GPUs meet each Unsloth minimum, see Unsloth minimums compared with catalog GPUs. The full copied table is in Unsloth QLoRA and LoRA VRAM minimums, and inference at other weight quants is in Q4 vs Q5 vs Q6 vs Q8.
Inference numbers are estimates from model configs and published file sizes, not measurements; training memory and speed weren’t measured here. How we calculate.
FAQ
Is fine-tuning with QLoRA more expensive than running the model?
Not always, going by Unsloth's minimums. For a 7B, Unsloth lists 5 GB for QLoRA, about the same as running a Q4 7B at 8K context (4.92 to 5.42 GiB). For 32B, its QLoRA minimum of 26 GB is above our 20.53 to 21.41 GiB inference estimate.
How much more does 16-bit LoRA need than QLoRA?
In Unsloth's table, roughly three to four times as much: 19 GB vs 5 GB at 7B, 33 GB vs 8.5 GB at 14B, and 76 GB vs 26 GB at 32B.
Did vramcalculator.io measure these fine-tuning numbers?
No. The QLoRA and LoRA figures are Unsloth's published minimums, copied on 22 October 2026. We didn't run the fine-tunes, and Unsloth doesn't state the sequence length behind its table.
Why is 14B inference above the 14B QLoRA minimum?
They measure different things. Our inference figure includes an 8K-token KV cache and runtime overhead on a specific GGUF file; Unsloth's is an absolute minimum for training that it says rises with batch size.
Numbers from calculator 0.3.0, model data 2026-10-02.1. Sources last checked 2026-10-29.