NeuroGenesysCan I run this LLM?
Last updated • By Will Schulz
An open-weight LLM needs memory for its weights, the KV cache for its context, and some working space. At 4-bit (Q4_K_M) Llama 3.1 8B needs about 7.13 GB with an 8K context and Llama 3.3 70B about 48.1 GB. Pick a model for its full breakdown, or a GPU or Mac to see everything that fits it.
How much memory does each model need?
| Model | 4-bit (Q4_K_M) | 16-bit (BF16) | Smallest GPU at 4-bit | Smallest Mac at 4-bit |
|---|---|---|---|---|
| Mistral 7B v0.3 | 6.64 GB | 17.0 GB | 8 GB | 24 GB |
| Qwen2.5 7B | 6.27 GB | 17.2 GB | 8 GB | 24 GB |
| Llama 3.1 8B | 7.13 GB | 18.6 GB | 8 GB | 24 GB |
| Qwen3 8B | 7.37 GB | 19.1 GB | 8 GB | 24 GB |
| Gemma 2 9B | 8.93 GB | 22.1 GB | 12 GB | 24 GB |
| Gemma 3 12B | 9.55 GB | 27.0 GB | 12 GB | 24 GB |
| Qwen2.5 14B | 11.9 GB | 32.9 GB | 12 GB | 24 GB |
| Phi-4 (14B) | 11.9 GB | 33.0 GB | 12 GB | 24 GB |
| gpt-oss-20b | 14.4 GB | 44.5 GB | 16 GB | 24 GB |
| Gemma 2 27B | 20.4 GB | 59.3 GB | 24 GB | 32 GB |
| Gemma 3 27B | 19.4 GB | 58.6 GB | 24 GB | 32 GB |
| Qwen3 30B-A3B | 21.0 GB | 64.6 GB | 24 GB | 32 GB |
| Qwen2.5 32B | 23.6 GB | 70.1 GB | 24 GB | 32 GB |
| Qwen3 32B | 23.8 GB | 70.7 GB | 24 GB | 32 GB |
| Mixtral 8x7B | 31.5 GB | 98.3 GB | 32 GB | 48 GB |
| Llama 3.1 70B | 48.1 GB | 149 GB | 48 GB | 64 GB |
| Llama 3.3 70B | 48.1 GB | 149 GB | 48 GB | 64 GB |
| Qwen2.5 72B | 49.5 GB | 153 GB | 80 GB | 64 GB |
| Llama 4 Scout | 71.2 GB | 227 GB | 80 GB | 96 GB |
| gpt-oss-120b | 75.0 GB | 242 GB | 80 GB | 96 GB |
| Qwen3 235B-A22B | 151 GB | 487 GB | 180 GB | 192 GB |
| Llama 4 Maverick | 254 GB | 827 GB | multi-GPU | 512 GB |
| Llama 3.1 405B | 261 GB | 841 GB | multi-GPU | 512 GB |
| DeepSeek-V3 / R1 | 424 GB | 1384 GB | multi-GPU | none |
Weights + FP16 KV cache + runtime overhead, with the VRAM calculator's defaults. Model and hardware data last checked Nov 09, 2026.
