nikimori/Qwen3.5-2B-Base-4bit

Qwen3.5-2B-Base, 4-bit MLX

Qwen/Qwen3.5-2B-Base converted to MLX with mlx_lm.convert -q --q-bits 4 --q-group-size 64 (mlx-lm 0.31.3, affine). The weights are otherwise unchanged, and the license is the upstream one (Apache-2.0).

Converted for Pretype, where it is the completion model for 8 GB Macs. On Pretype's 5,649-row, 17-language autocomplete set it gets the first word right on 17.1% of all rows, against 16.1% for the post-trained mlx-community/Qwen3.5-2B-4bit at the same size (paired McNemar p = 0.005).

from mlx_lm import load, generate
model, tokenizer = load("<this repo>")
print(generate(model, tokenizer, prompt="The quick brown fox", max_tokens=20))