GELab-Zero-4B-preview-Sico-Evolution-MLX-4bit

GELab-Zero-4B-preview-Sico-Evolution-MLX-4bit

4-bit MLX quantization of microsoft/GELab-Zero-4B-preview-Sico-Evolution — Microsoft's Qwen3-VL-based GUI agent (Edge / Copilot UI trajectories, from the Sico evolution pipeline) — running natively on Apple Silicon via mlx-vlm.

At a glance

Verification (2026-07-12, M-series Mac)

Converted with mlx-vlm 0.3.9 and image-inference smoke-tested locally before publishing — the BF16 source and this quantized pack both correctly describe the model's own benchmark chart:

"The image displays a bar chart comparing the Task Success Rate (TSR) of various models in Edge/Copilot test cases."

Sibling variants

<!-- sibling-table -->

Usage

Requires mlx-vlm >= 0.3.9:

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("majentik/GELab-Zero-4B-preview-Sico-Evolution-MLX-4bit")
prompt = apply_chat_template(processor, model.config, "Describe this screenshot.", num_images=1)
print(generate(model, processor, prompt, ["screenshot.png"], max_tokens=256))

python -m mlx_vlm generate \
    --model majentik/GELab-Zero-4B-preview-Sico-Evolution-MLX-4bit \
    --image screenshot.png --prompt "What action should I take next?"

Intended use

GUI-agent tasks — screenshot understanding, UI grounding, and next-action prediction for desktop/web automation — locally on Apple Silicon. This is the Edge/Copilot-evolved variant; see the base model card for benchmarks (82.9% TSR) and the Sico pipeline.

Reproduce

python -m mlx_vlm convert --hf-path microsoft/GELab-Zero-4B-preview-Sico-Evolution \
    --mlx-path GELab-4bit -q --q-bits 4 --q-group-size 64

License

Apache-2.0, inherited from the upstream model. This repository redistributes quantized weights with attribution; training and evaluation details are on the base model card.