Apple Silicon RAM Calculator
Can your Mac run this model? Select a model and quantization to see which chips have enough RAM — and how fast they'll run it.
Check a model
Quick reference: RAM by model size
Approximate VRAM/unified memory needed at each quantization. Add ~2–4 GB for OS and runtime overhead.
| Model size | Q4_K_M | Q5_K_M | Q6_K | Q8_0 | Minimum Mac |
|---|---|---|---|---|---|
| 1B | ~0.8 GB | ~1.0 GB | ~1.2 GB | ~1.5 GB | 8 GB (any M-series) |
| 3B | ~2.0 GB | ~2.5 GB | ~3.0 GB | ~3.5 GB | 8 GB (any M-series) |
| 7B–8B | ~4.5 GB | ~5.5 GB | ~6.5 GB | ~8.5 GB | 16 GB (M-series base) |
| 14B | ~9 GB | ~11 GB | ~13 GB | ~16 GB | 24 GB (M Pro+) |
| 32B | ~20 GB | ~24 GB | ~29 GB | ~35 GB | 36–48 GB (M Max) |
| 70B | ~43 GB | ~53 GB | ~63 GB | ~75 GB | 64 GB (M Max 64 GB+) |
| 105B | ~65 GB | ~79 GB | ~94 GB | ~112 GB | 128 GB (M Max 128 GB) |
| 235B (MoE) | ~130–140 GB | ~160 GB | ~190 GB | ~240 GB | 192 GB (M Ultra) |
| 405B | ~245 GB | ~300 GB | — | — | 512 GB (M3 Ultra) |
MoE (Mixture of Experts) models like Qwen 3 235B A22B use fewer active parameters during inference — they need less RAM than their total parameter count suggests. A 235B MoE model at Q4 needs ~130–140 GB, not ~145 GB.
Apple Silicon RAM tiers
| Chip | RAM options | Largest model at Q4_K_M | Best for |
|---|---|---|---|
| M4, M3, M2, M1 (base) | 8–32 GB | 8B (16 GB) · 14B (24 GB) | 7B–8B daily use |
| M4 Pro, M3 Pro, M2 Pro | 24–64 GB | 14B (24 GB) · 32B (48+ GB) | 14B daily, occasional 32B |
| M4 Max, M3 Max, M2 Max | 36–128 GB | 32B (48 GB) · 70B (128 GB) | 32B–70B inference |
| M2 Ultra, M3 Ultra | 64–512 GB | 235B (192+ GB) · 405B (512 GB) | Maximum model size |
About quantization and quality
| Quantization | Size vs F32 | Quality | Speed | Recommended use |
|---|---|---|---|---|
| Q2_K | ~25% | Noticeably degraded | Fastest | When RAM is severely limited |
| Q3_K_M | ~35% | Somewhat degraded | Very fast | When RAM is tight |
| Q4_K_M | ~45% | Good — minimal loss | Fast | Best daily driver |
| Q5_K_M | ~55% | Very good | Moderately fast | Quality-focused use |
| Q6_K | ~65% | Excellent — near full | Moderate | High-quality tasks |
| Q8_0 | ~83% | Near-lossless | Slower | Benchmarking, max quality |
Related tools and guides
Data
benchmarks.json — full dataset · chips.json — chip summaries · benchmarks.csv — CSV export