Minimum RAM to Run a 70B LLM on Apple Silicon
How much unified memory do you actually need to run Llama 3.3 70B or Qwen 3 70B on a Mac? Here is the measured answer.
The RAM math for 70B models
A 70B parameter model at Q4_K_M quantization occupies approximately 35–42 GB of unified memory. At Q5_K_M it rises to ~46 GB; at Q8_0 it reaches ~70 GB. These numbers come from measured weight files, not estimates.
Apple Silicon needs headroom beyond the weights themselves: the OS, the KV cache (which grows with context length), and runtime overhead. A 64 GB chip running a 42 GB Q4 model has ~20 GB headroom — enough for a reasonable context window. A 48 GB chip can technically load the weights but has very little room for the KV cache, and may behave erratically at longer contexts.
48 GB is technically possible at Q3_K_S quantization (~28 GB) but quality degrades significantly. 64 GB gives you comfortable headroom for Q4–Q5 quantization at 70B scale. For Q8 quality, you need 96 GB minimum, 128 GB preferred.
Published 70B benchmark results
All published results for 70B-class models from the SiliconBench dataset. Sorted by tok/s descending.
| Chip | Model | Quant | Avg tok/s | Source | |
|---|---|---|---|---|---|
| M4 Max (24-core GPU) | Llama 3.3 70B | Q5_K_M | 50 GB | 7.1 tok/s | ref |
More 70B results are being aggregated. Follow the feed for updates.
Which chips can run 70B?
Can run 70B (Q4–Q5)
- M4 Max (64 GB) — 7.1+ tok/s measured
- M4 Max (128 GB) — comfortable headroom
- M3 Max (64 GB) — tight but functional at Q4
- M3 Max (96 GB+) — comfortable
- M2 Max (96 GB) — works but slower generation
- M1 Ultra (128 GB) — works
Cannot run 70B comfortably
- M4 Pro (48 GB max) — 70B at Q3 only; quality loss
- M3 Pro (36 GB max) — too small for 70B
- M4 (32 GB max) — 70B does not fit
- M3 (24 GB max) — 70B does not fit
- Any chip with less than 48 GB — too small
Related
Data
benchmarks.json — full dataset · models.json — model summaries