Yes, just — 5.2 GB spare
Llama 3.3 70B Instruct at Q4_K_M fits your M3 Max · 64 GB entirely in unified memory at 8K context, at an estimated 5.6 tokens per second. Past 24K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully in unified memory
8K context
Q4_K_M · 39.7 GB
Llama 3.3 Community
Released Dec 2024
Still the creative-writing favourite: consistent voice, takes direction. Needs 48 GB to sit comfortably on GPU at Q4.
The VRAM budget
weights 39.7 GB
Weights 39.7 GB
KV cache @ 8K 2.50 GB
Runtime overhead 0.6 GB
Free 5.2 GB of 48.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 69.9 GB | 73.0 GB | — | ~3.2 | −0.1% ppl | 25.0 GB over |
| Q6_K | 53.9 GB | 57.0 GB | — | ~4.1 | −0.4% ppl | 9.0 GB over |
| Q5_K_M | 46.6 GB | 49.7 GB | 2K | ~4.8 | −0.8% ppl | 1.7 GB over |
| Q4_K_M | 39.7 GB | 42.8 GB | 24K | 5.6 | −1.9% ppl | Recommended |
| Q3_K_M | 32.1 GB | 35.2 GB | 48K | 7.0 | −5.4% ppl | Long context |
| Q2_K | 27.5 GB | 30.6 GB | 63K | 8.1 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/Llama-3.3-70B-Instruct-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 39.7 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02macOS caps what the GPU may wire down at about 48.0 GB of 64 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03Only 5.2 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 24K context.