How much memory do popular local AI models need?
Estimated memory at Q4_K_M with an 8K context, for the most popular open models on Hugging Face. Pick a model for the full table by quantization and context, or compare them all in the model fit chart.
| Model | Parameters | Type | Memory, Q4_K_M, 8K | Smallest Mac |
|---|---|---|---|---|
| Gemma 4 12B | 11.96B | Dense | 9.2 GB | 16 GB |
| gpt-oss 20B | 20.91B | MoE, 32 experts | 14 GB | 24 GB |
| Gemma 4 26B-A4B | 25.81B | MoE, 128 experts | 17 GB | 24 GB |
| Qwen3.8 27B | 27.78B | Dense | 19 GB | 32 GB |
| GLM-4.7 Flash | 31.22B | MoE, 64 experts | 21 GB | 32 GB |
| Gemma 4 31B | 31.27B | Dense | 22 GB | 32 GB |
| Qwen3.6 35B-A3B | 35.95B | MoE, 256 experts | 23 GB | 32 GB |
| gpt-oss 120B | 117B | MoE, 128 experts | 73 GB | 128 GB |
| Mistral Medium 3.5 | 128B | Dense | 82 GB | 128 GB |
| Qwen3.8 Flash Next | 180B | MoE, 512 experts | 111 GB | 192 GB |
| DeepSeek V4 Flash | 291B | MoE, 256 experts | 179 GB | 256 GB |
| GLM-5.3 | 753B | MoE, 256 experts | 462 GB | None |
| DeepSeek V4 Pro | 1,599B | MoE, 384 experts | 978 GB | None |
| Kimi K3 | 2,780B | MoE, 896 experts | 1,700 GB | None |
Parameters and attention shapes read from each model's files on Hugging Face on 2026-11-09. Estimates include the weights, the KV cache and 1 GB of buffers, against about three quarters of a Mac's unified memory. How the estimate works.