Local LLM VRAM requirements
How much GPU or Mac memory each open model needs. Estimates from official parameter counts.
Quick answer
Rule of thumb: at Q4 a model needs about 0.6 GB per billion parameters, plus about 1.5 GB headroom. A 7B to 8B model fits in 8 GB, a 14B model in 12 GB, a 27B to 32B model in 24 GB, and a 70B model needs about 44 GB.
| Model | Params | Q4 needs |
|---|---|---|
| Ministral 3 3B | 3B | 3.3 GB |
| Phi-4 mini 3.8B | 3.8B | 3.8 GB |
| Qwen3 4B | 4B | 3.9 GB |
| Gemma 4 E4B | 8B | 6.3 GB |
| Ministral 3 8B | 8B | 6.3 GB |
| Llama 3.1 8B Instruct | 8B | 6.3 GB |
| Qwen3 8B | 8.2B | 6.4 GB |
| Gemma 4 12B | 11.95B | 8.7 GB |
| Ministral 3 14B | 14B | 9.9 GB |
| Phi-4 14B | 14.7B | 10 GB |
| DeepSeek R1 Distill Qwen 14B | 14.8B | 10 GB |
| gpt-oss-20b | 21B | 14 GB |
| Gemma 4 26B A4B | 25.2B | 17 GB |
| Qwen3.6 27B | 27B | 18 GB |
| Qwen3.8 27B | 27.8B | 18 GB |
| Muse Glimmer 30B | 29.6B | 19 GB |
| Qwen3 Coder 30B A3B | 30.5B | 20 GB |
| Gemma 4 31B | 30.7B | 20 GB |
| DeepSeek R1 Distill Qwen 32B | 32.8B | 21 GB |
| Qwen3.6 35B A3B | 35B | 22 GB |
| Llama 3.3 70B Instruct | 70B | 44 GB |
| gpt-oss-120b | 117B | 72 GB |
| Mistral Small 4 | 119B | 73 GB |
| Qwen3.8 Flash Next | 180B | 110 GB |
| Muse Spark Open | 816B | 491 GB |
Memory needed = weights at that quant + 1.5 GB. Mixture-of-experts models still need memory for all parameters.
FAQ
How much VRAM do I need for a 70B model?
About 44 GB at Q4, so two 24GB GPUs or a Mac with 64GB or more unified memory.
How much VRAM do I need for a 7B or 8B model?
About 6.3 GB at Q4, so an 8GB GPU works.
Does a mixture-of-experts model need less VRAM?
No. It needs memory for all parameters, but it runs faster because only the active experts are read per token.