LLMs by VRAM Size
Choose how much GPU memory you have. 8 GB runs 7–8B models, 12–16 GB runs 14B models, 24 GB runs 32B models, and 48 GB runs 70B models at Q4_K_M.
| VRAM | Models that fit | Largest |
|---|---|---|
| 4 GB | 7 of 58 | Gemma 4 E2B |
| 6 GB | 13 of 58 | Gemma 4 E4B |
| 8 GB | 18 of 58 | Qwen 3.5 9B |
| 10 GB | 21 of 58 | Ministral 3 14B |
| 12 GB | 25 of 58 | Qwen 3 14B |
| 16 GB | 28 of 58 | Gemma 4 26B-A4B |
| 20 GB | 35 of 58 | Gemma 4 31B |
| 24 GB | 41 of 58 | Qwen 3.5 35B-A3B |
| 32 GB | 42 of 58 | Mixtral 8x7B |
| 48 GB | 46 of 58 | Qwen3-Next 80B-A3B |
| 64 GB | 48 of 58 | gpt-oss-120b |
| 80 GB | 50 of 58 | Qwen 3.5 122B-A10B |
| 96 GB | 50 of 58 | Qwen 3.5 122B-A10B |
| 128 GB | 51 of 58 | Qwen 3.5 122B-A10B |
How to use the VRAM tiers
Each tier page lists every model we track that fits fully in that much memory at Q4_K_M with an 8K context, the best quantization each model can use, and the longest context that still fits. Models that only fit at a lower quantization are listed separately, so you can decide whether the quality trade-off is worth it. Each page also shows which image and video models fit, and which GPUs and Macs have that much memory.
We keep a 6% safety margin, because the desktop, browser and driver use some memory on most systems. A model marked as a tight fit usually works on a GPU that is not driving a display, or with a slightly shorter context.
Rules of thumb
- 4–6 GB: small 1–4B models, or 7–8B models with CPU offloading.
- 8 GB: 7–9B models at Q4_K_M with a moderate context.
- 12 GB: 12–14B models; FLUX.1 with GGUF weights.
- 16 GB: 14B at high quality, gpt-oss-20b, 24B models with short context.
- 24 GB: 27–35B models, the sweet spot for a single consumer GPU.
- 32 GB: the same models at Q6_K or with much longer context.
- 48 GB: 70B models at Q4_K_M, usually two 24 GB GPUs.
- 64–128 GB: 70B at Q8_0 and 80–120B mixture-of-experts models on Macs and unified-memory PCs.
Not sure which tier you are in? Use Can I Run It with your exact GPU, or read the best local LLMs for every VRAM size for our picks.