Qwen2.5-Coder 14B on a Mac
Alibaba · 14.8B parameters · Apache 2.0 · released 2024-11
Weights @ Q4_K_M
8.4 GiB
the usual download size
KV cache
192 KB
per token, FP16
Max context
128K
tokens
Smallest Mac
16 GB
M1
Short answerA M1 16GB is the smallest Apple Silicon machine that loads Qwen2.5-Coder 14B at Q4_K_M with an 8K context, at roughly 6 tokens/sec. MacBook Air (M1) is the cheapest way to get that configuration.
Every Mac, ranked
| Mac | Memory | Needs | Tok/s | Verdict |
|---|---|---|---|---|
| M1 8GB | 8 GB | 11.2 GiB | 6 | No |
| M1 16GB | 16 GB | 11.2 GiB | 6 | Yes, but tight |
| M2 16GB | 16 GB | 11.2 GiB | 9 | Yes, but tight |
| M4 16GB | 16 GB | 11.2 GiB | 11 | Yes, but tight |
| M1 Pro 16GB | 16 GB | 11.2 GiB | 18 | Yes, but tight |
| M3 Pro 18GB | 18 GB | 11.2 GiB | 13 | Yes, comfortably |
| M2 24GB | 24 GB | 11.2 GiB | 9 | Yes, comfortably |
| M4 24GB | 24 GB | 11.2 GiB | 11 | Yes, comfortably |
| M4 Pro 24GB | 24 GB | 11.2 GiB | 24 | Yes, comfortably |
| M4 32GB | 32 GB | 11.2 GiB | 11 | Yes, comfortably |
| M1 Max 32GB | 32 GB | 11.2 GiB | 35 | Yes, comfortably |
| M2 Max 32GB | 32 GB | 11.2 GiB | 35 | Yes, comfortably |
| M3 Max 36GB | 36 GB | 11.2 GiB | 26 | Yes, comfortably |
| M4 Max 36GB | 36 GB | 11.2 GiB | 36 | Yes, comfortably |
| M4 Pro 48GB | 48 GB | 11.2 GiB | 24 | Yes, comfortably |
| M4 Max 48GB | 48 GB | 11.2 GiB | 48 | Yes, comfortably |
| M4 Pro 64GB | 64 GB | 11.2 GiB | 24 | Yes, comfortably |
| M1 Max 64GB | 64 GB | 11.2 GiB | 35 | Yes, comfortably |
| M4 Max 64GB | 64 GB | 11.2 GiB | 48 | Yes, comfortably |
| M2 Ultra 64GB | 64 GB | 11.2 GiB | 71 | Yes, comfortably |
| M2 Max 96GB | 96 GB | 11.2 GiB | 35 | Yes, comfortably |
| M3 Ultra 96GB | 96 GB | 11.2 GiB | 72 | Yes, comfortably |
| M3 Max 128GB | 128 GB | 11.2 GiB | 35 | Yes, comfortably |
| M4 Max 128GB | 128 GB | 11.2 GiB | 48 | Yes, comfortably |
| M2 Ultra 192GB | 192 GB | 11.2 GiB | 71 | Yes, comfortably |
| M3 Ultra 256GB | 256 GB | 11.2 GiB | 72 | Yes, comfortably |
| M3 Ultra 512GB | 512 GB | 11.2 GiB | 72 | Yes, comfortably |
Download size by quantisation
Weight memory scales linearly with bits per parameter. Everything below assumes an 8K context on top.
| Quantisation | Weights | Total @ 8K | Notes |
|---|---|---|---|
| FP16 | 27.6 GiB | 31.2 GiB | Full precision. Reference quality, twice the memory of Q8. |
| Q8_0 | 14.6 GiB | 17.7 GiB | Indistinguishable from FP16 in practice, at half the size. |
| Q6_K | 11.4 GiB | 14.2 GiB | Near-lossless. A good stop when you have memory to spare. |
| Q5_K_M | 9.8 GiB | 12.6 GiB | Slightly better than Q4_K_M, noticeably bigger. |
| Q4_K_M | 8.4 GiB | 11.2 GiB | The default. Best quality-per-gigabyte for most people. |
| Q3_K_M | 6.7 GiB | 9.4 GiB | Visible quality loss. Use to squeeze one size class up. |
| Q2_K | 5.2 GiB | 7.7 GiB | Last resort. Often worse than a smaller model at Q4. |