Run a Local LLM Without a Dedicated GPU
Find GGUF models for CPU-only PCs by system RAM and for Apple silicon Macs by unified memory, with quantization levels and 4K memory estimates.
Memory estimates at 4,096 tokens for one text sequence.
CPU only · system RAM
Use llama.cpp to load GGUF on the CPU. Leave RAM for the OS and open applications. Generation speed also depends on the processor and memory bandwidth.
| Memory | Example chat model |
|---|---|
| 16 GB | Xing4.0 29B-A4B GGUF · Q2_K · 13.6 GiB |
| 32 GB | Kolibri-1 78B-A3.5B GGUF · IQ2_XS · 25.6 GiB |
| 64 GB | Llama 4 Scout 17B (16E) GGUF · Q3_K_M · 55.9 GiB |
| 128 GB | Qwen3.8-Flash-Next GGUF · Q4_K_M · 79.3 GiB |
Apple silicon unified memory
The Mac GPU shares memory with the system. The calculator budgets 75% of installed memory before checking headroom.
| Memory | Example chat model |
|---|---|
| 16 GB | Hy-Memory-30B-A3B GGUF · Q8_0-expsIQ2_XXS · 9.9 GiB |
| 24 GB | Xing4.0 29B-A4B GGUF · IQ3_XXS · 15.3 GiB |
| 36 GB | Kolibri-1 78B-A3.5B GGUF · IQ1_M · 20.9 GiB |
| 48 GB | Kolibri-1 78B-A3.5B GGUF · IQ2_XS · 25.6 GiB |
| 64 GB | GLM-4.5-Air GGUF · Q2_K · 37.2 GiB |
| 128 GB | Qwen3.8-Flash-Next GGUF · Q4_K_M · 79.3 GiB |