Home → LLM VRAM
LLM VRAM requirements by model size
These are model file sizes from the Ollama library (retrieved October 2026). File size is the dominant part of the VRAM floor; expect runtime usage to land slightly above it once context and activations are counted.
DeepSeek-R1
| Variant | Quantization | File size |
|---|---|---|
| 1.5B | — | 1.1GB |
| 7B | Q4 | 4.7GB |
| 7B | Q8 | 8.1GB |
| 7B | FP16 | 15GB |
| 8B | Q4 | 5.2GB |
| 14B | Q4 | 9.0GB |
| 14B | Q8 | 16GB |
| 14B | FP16 | 30GB |
| 32B | Q4 | 20GB |
| 32B | Q8 | 35GB |
| 32B | FP16 | 66GB |
| 70B | Q4 | 43GB |
| 70B | Q8 | 75GB |
| 70B | FP16 | 141GB |
| 671B | Q4 | 404GB |
Llama 3.1
| Variant | Quantization | File size |
|---|---|---|
| 8B | Q4 | 4.7GB |
| 8B | Q8 | 8.5GB |
| 8B | FP16 | 16GB |
| 70B | Q4 | 43GB |
| 405B | — | 243GB |
Qwen 2.5
| Variant | Quantization | File size |
|---|---|---|
| 7B | Q4 | 4.7GB |
| 14B | Q4 | 9.0GB |
| 32B | Q4 | 20GB |
| 72B | Q4 | 47GB |
Source: ollama.com/library model pages (deepseek-r1, llama3.1, qwen2.5), retrieved Oct 2026.
- One 24GB card (RTX 3090 / 4090 class) runs up to ~32B models at Q4 — the 20GB file plus overhead.
- 70B at Q4 (43GB file) needs 2×24GB or a 48GB+ card.
- 8B–14B at Q4 fits on a single 12GB card.
Reading the quantization ladder
Each step up in precision roughly doubles the footprint: DeepSeek-R1 7B goes 4.7GB → 8.1GB → 15GB across Q4 → Q8 → FP16. Q4 is the common local-deployment default because it roughly quarters the FP16 footprint with modest quality loss on most tasks; if you have the memory, Q8 is a low-risk upgrade.
If the model you want exceeds what your card holds, the standard alternatives are a smaller quant, a smaller variant, or renting a larger card for the session.