aufklarer/Qwen3-4B-Instruct-2507-MLX-4bit

🤗 Hugging Face sourcetext-generationapache-2.04B params8.0 GBsafetensors✓ 2 checksumsupdated today

Submit in one command

Run it next to your model folder. It makes the torrent, checks your files against Hugging Face, and submits it. You just start seeding and paste your key from your account. It only reads your files and never changes them. Read the script first if you like.

curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo aufklarer/Qwen3-4B-Instruct-2507-MLX-4bit ./model-folder

Needs a seeder →

Qwen3-4B-Instruct-2507 — MLX int4

First-party MLX export of Qwen/Qwen3-4B-Instruct-2507, quantized to int4 (group size 64) for on-device chat on Apple Silicon. Built by our own pipeline (speech-models/export_mlx.py, via mlx_lm.convert).

Runs in the runner voice companion through a hand-written MLX dense runtime (soniqo/speech-swift → Qwen3Chat/Qwen3DenseModel), not a generic loader — the forward pass is numerically parity-verified against mlx_lm (identical next-token logits).


Params	4B (dense) · 36 layers · 32 q / 8 kv heads · head_dim 128
Quantization	int4, group size 64 (~4.5 bits/weight, 2.28 GB)
Context	262144

Attribution & license

Weights: derivative of Qwen/Qwen3-4B-Instruct-2507, Alibaba/Qwen — Apache-2.0.
Conversion: mlx_lm.convert (Apple MLX) — MIT.