Yes — with 7.2 GB to spare
Qwen3.5 4B at Q4_K_M fits your M1 · 16 GB entirely in unified memory at 8K context, at an estimated 15 tokens per second. Past 239K 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 · 2.6 GB
Apache 2.0
Released 28 Feb 2026
Vision
The 8 GB coding agent. Q4 lands near 3.4 GB, leaving room for a real context window.
The VRAM budget
weights 2.6 GB
Weights 2.6 GB
KV cache @ 8K 0.25 GB
Runtime overhead 0.6 GB
Free 7.2 GB of 10.7 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 8.7 GB | 9.5 GB | 45K | 4.4 | Reference | Long context |
| Q8_0 | 4.6 GB | 5.5 GB | 175K | 8.2 | −0.1% ppl | Long context |
| Q6_K | 3.6 GB | 4.4 GB | 209K | 11 | −0.4% ppl | Long context |
| Q5_K_M | 3.1 GB | 3.9 GB | 224K | 12 | −0.8% ppl | Long context |
| Q4_K_M | 2.6 GB | 3.5 GB | 239K | 15 | −1.9% ppl | Recommended |
| Q3_K_M | 2.1 GB | 3.0 GB | 255K | 18 | −5.4% ppl | Long context |
| Q2_K | 1.8 GB | 2.7 GB | 256K | 21 | −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. Only 8 of its 32 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/Qwen3.5-4B-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 2.6 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 10.7 GB of 16 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to 239K context on this card.