SuperQwen-AgentWorld-35B-A3B-abliterated-mlx-4bit
on a Mac: what it actually needs
For Jun Song. Worked out from your repo, no sign-up,
nothing to click.
4-bit
40 layers
2 KV heads
head dim 256
7 likes
191 downloads
19.5 GB
measureddownload size, summed from the
4 safetensors files in your repo
32 GB Mac
computedthe smallest machine that holds it
at 8k context, assuming MLX can wire 75% of unified memory
What it needs as context grows
4k context0.34 GB KV + 19.5 GB weights + 1.2 GB runtime = 21.0 GB, smallest Mac 32 GB
8k context0.67 GB KV + 19.5 GB weights + 1.2 GB runtime = 21.4 GB, smallest Mac 32 GB
32k context2.68 GB KV + 19.5 GB weights + 1.2 GB runtime = 23.4 GB, smallest Mac 32 GB
128k context10.74 GB KV + 19.5 GB weights + 1.2 GB runtime = 31.4 GB, smallest Mac 48 GB
KV cache = 2 tensors x 40 layers x 2 KV heads
x 256 head dim x 2 bytes (fp16) x tokens. Read straight off your
config.json. Runtime overhead is the 1.2 GB I see for Python, MLX, the
tokenizer and activations on my own runs.
Mac by Mac
| Context | Needs | 8 | 16 | 18 | 24 | 32 | 36 | 48 | 64 | 96 | 128 |
| 4k | 21.0 GB | no | no | no | no | yes | yes | yes | yes | yes | yes |
| 8k | 21.4 GB | no | no | no | no | yes | yes | yes | yes | yes | yes |
| 32k | 23.4 GB | no | no | no | no | tight | yes | yes | yes | yes | yes |
| 128k | 31.4 GB | no | no | no | no | no | no | yes | yes | yes | yes |
Columns are unified memory in GB. "Tight" means it fits but leaves under
2 GB for anything else, so expect swap once a browser is open.
Run it
pip install mlx-lm
mlx_lm.generate --model Jiunsong/SuperQwen-AgentWorld-35B-A3B-abliterated-mlx-4bit \
--prompt "hello" --max-tokens 100
What is not on this page
not measuredTokens per second. That needs a
real run on real silicon and I will not print a number I have not seen. If you want it,
reply and I will run SuperQwen-AgentWorld-35B-A3B-abliterated-mlx-4bit on an M5 and add the result here with the
command I used.