Yes — with 68.0 GB to spare
Qwen3 4B at Q4_K_M fits your M2 Max · 96 GB entirely in unified memory at 8K context, at an estimated 99 tokens per second. There is room for its full 32K window.
Fully in unified memory
8K context
Q4_K_M · 2.3 GB
Apache 2.0
Released Apr 2025
The 2025 sweet spot for 8 GB cards with reasoning traces. Qwen3.5 4B adds vision and 8× the context.
The VRAM budget
weights 2.3 GB
Weights 2.3 GB
KV cache @ 8K 1.13 GB
Runtime overhead 0.6 GB
Free 68.0 GB of 72.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 7.5 GB | 9.2 GB | 32K | 30 | Reference | Long context |
| Q8_0 | 4.0 GB | 5.7 GB | 32K | 56 | −0.1% ppl | Long context |
| Q6_K | 3.1 GB | 4.8 GB | 32K | 73 | −0.4% ppl | Long context |
| Q5_K_M | 2.7 GB | 4.4 GB | 32K | 84 | −0.8% ppl | Long context |
| Q4_K_M | 2.3 GB | 4.0 GB | 32K | 99 | −1.9% ppl | Recommended |
| Q3_K_M | 1.8 GB | 3.6 GB | 32K | 122 | −5.4% ppl | Long context |
| Q2_K | 1.6 GB | 3.3 GB | 32K | 143 | −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/Qwen3-4B-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 2.3 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 32K context on this card.