qwen3-4b-Z-Image-Engineer-V2-6bit-MLX

tokenfires/qwen3-4b-Z-Image-Engineer-V2-6bit-MLX

Mirror of BennyDaBall/qwen3-4b-Z-Image-Engineer-V2-6bit-MLX, a Qwen3-4B fine-tune that expands short prompts into detailed Z-Image-Turbo prompts. Kept under tokenfires so a load-bearing local model can be restored if the upstream repo changes or disappears.

Quantization: affine, 6 bits, group size 64. The weights are unchanged from upstream. See the upstream card for usage notes and the recommended system prompt (system_prompt.json is included here).

Use with LM Studio

Search for tokenfires/qwen3-4b-Z-Image-Engineer-V2-6bit-MLX in the LM Studio model downloader, or open this page and choose Use this model → LM Studio.

Use with mlx-lm

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("tokenfires/qwen3-4b-Z-Image-Engineer-V2-6bit-MLX")

messages = [{"role": "user", "content": "a lighthouse at dusk"}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, verbose=True)