Yes — with 65.7 GB to spare
Qwen3 8B at Q4_K_M fits your M2 Max · 96 GB entirely in unified memory at 8K context, at an estimated 49 tokens per second. There is room for its full 128K window.
Fully in unified memory
8K context
Q4_K_M · 4.6 GB
Apache 2.0
Released Apr 2025
Apache-2.0 alternative to Llama 3.1 8B, with a switchable thinking mode.
The VRAM budget
weights 4.6 GB
Weights 4.6 GB
KV cache @ 8K 1.13 GB
Runtime overhead 0.6 GB
Free 65.7 GB of 72.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 15.3 GB | 17.0 GB | 128K | 15 | Reference | Long context |
| Q8_0 | 8.1 GB | 9.8 GB | 128K | 28 | −0.1% ppl | Long context |
| Q6_K | 6.3 GB | 8.0 GB | 128K | 36 | −0.4% ppl | Long context |
| Q5_K_M | 5.4 GB | 7.1 GB | 128K | 41 | −0.8% ppl | Long context |
| Q4_K_M | 4.6 GB | 6.3 GB | 128K | 49 | −1.9% ppl | Recommended |
| Q3_K_M | 3.7 GB | 5.5 GB | 128K | 60 | −5.4% ppl | Long context |
| Q2_K | 3.2 GB | 4.9 GB | 128K | 70 | −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/Qwen3-8B-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 4.6 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 72.0 GB of 96 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.