Yes — with 14.9 GB to spare
Granite 4.1 3B at Q4_K_M fits your M3 · 24 GB entirely in unified memory at 8K context, at an estimated 29 tokens per second. There is room for its full 128K window.
Fully in unified memory
8K context
Q4_K_M · 1.9 GB
Apache 2.0
Released 29 Apr 2026
Dense, small, enterprise-flavoured: tool calling and instruction following, no thinking mode.
The VRAM budget
weights 1.9 GB
Weights 1.9 GB
KV cache @ 8K 0.63 GB
Runtime overhead 0.6 GB
Free 14.9 GB of 18.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 6.3 GB | 7.6 GB | 128K | 8.8 | Reference | Long context |
| Q8_0 | 3.4 GB | 4.6 GB | 128K | 17 | −0.1% ppl | Long context |
| Q6_K | 2.6 GB | 3.8 GB | 128K | 22 | −0.4% ppl | Long context |
| Q5_K_M | 2.2 GB | 3.5 GB | 128K | 25 | −0.8% ppl | Long context |
| Q4_K_M | 1.9 GB | 3.1 GB | 128K | 29 | −1.9% ppl | Recommended |
| Q3_K_M | 1.5 GB | 2.8 GB | 128K | 36 | −5.4% ppl | Long context |
| Q2_K | 1.3 GB | 2.6 GB | 128K | 42 | −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/granite-4.1-3b-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 1.9 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 128K context on this card.