Yes — with 380.7 GB to spare
Llama 3.2 3B Instruct at Q4_K_M fits your M3 Ultra · 512 GB entirely in unified memory at 8K context, at an estimated 248 tokens per second. There is room for its full 128K window.
Fully in unified memory
8K context
Q4_K_M · 1.8 GB
Llama 3.2 Community
Released Sep 2024
The 2024 "it just runs" model for 8 GB laptops. Qwen3.5 4B does the same job better now.
The VRAM budget
weights 1.8 GB
Weights 1.8 GB
KV cache @ 8K 0.88 GB
Runtime overhead 0.6 GB
Free 380.7 GB of 384.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 6.0 GB | 7.5 GB | 128K | 75 | Reference | Long context |
| Q8_0 | 3.2 GB | 4.7 GB | 128K | 141 | −0.1% ppl | Long context |
| Q6_K | 2.5 GB | 3.9 GB | 128K | 182 | −0.4% ppl | Long context |
| Q5_K_M | 2.1 GB | 3.6 GB | 128K | 211 | −0.8% ppl | Long context |
| Q4_K_M | 1.8 GB | 3.3 GB | 128K | 248 | −1.9% ppl | Recommended |
| Q3_K_M | 1.5 GB | 2.9 GB | 128K | 306 | −5.4% ppl | Long context |
| Q2_K | 1.3 GB | 2.7 GB | 128K | 357 | −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-3.2-3B-Instruct-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 1.8 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 384.0 GB of 512 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to the model's full 128K context on this card.