Yes — with 16.2 GB to spare
Qwen3 0.6B at Q4_K_M fits your M2 · 24 GB entirely in unified memory at 8K context, at an estimated 166 tokens per second. There is room for its full 32K window.
Fully in unified memory
8K context
Q4_K_M · 0.3 GB
Apache 2.0
Released Apr 2025
Useful mostly as a speculative-decoding draft model for its larger siblings.
The VRAM budget
weights 0.3 GB
Weights 0.3 GB
KV cache @ 8K 0.88 GB
Runtime overhead 0.6 GB
Free 16.2 GB of 18.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 1.1 GB | 2.6 GB | 32K | 50 | Reference | Long context |
| Q8_0 | 0.6 GB | 2.1 GB | 32K | 94 | −0.1% ppl | Long context |
| Q6_K | 0.5 GB | 1.9 GB | 32K | 122 | −0.4% ppl | Long context |
| Q5_K_M | 0.4 GB | 1.9 GB | 32K | 141 | −0.8% ppl | Long context |
| Q4_K_M | 0.3 GB | 1.8 GB | 32K | 166 | −1.9% ppl | Recommended |
| Q3_K_M | 0.3 GB | 1.7 GB | 32K | 205 | −5.4% ppl | Long context |
| Q2_K | 0.2 GB | 1.7 GB | 32K | 239 | −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/Qwen3-0.6B-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 0.3 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 18.0 GB of 24 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.