# Microsoft previews GELab-Zero-4B, a compact GUI agent

> The 4-billion-parameter vision-language model targets on-screen and mobile automation, built atop Qwen3-VL.

Published by The Open Weights on Jun 30, 2026. Canonical: https://theopenweights.com/news/gelab-zero-4b-f2ea

## Key facts

- Company: Microsoft
- Model: GELab-Zero-4B (preview)
- Version: Zero-4B
- Category: Vision-Language
- Modalities: Vision-Language
- License: Other (Open weights, commercial use allowed)
- Parameters: 4B
- Architecture: Qwen3VLForConditionalGeneration
- Precision: BF16
- Weights size: 8.9 GB
- Significance: minor
- Published: 2026-06-30
- Last verified: 2026-07-05
- Hugging Face: https://huggingface.co/microsoft/GELab-Zero-4B-preview-Sico-Evolution
- Canonical URL: https://theopenweights.com/news/gelab-zero-4b-f2ea

Microsoft has quietly posted a preview of **GELab-Zero-4B**, a compact vision-language model aimed at graphical user interface and mobile automation. According to the model's [Hugging Face page](https://huggingface.co/microsoft/GELab-Zero-4B-preview-Sico-Evolution), the roughly 4-billion-parameter model is built on Alibaba's Qwen3-VL foundation and framed as an agent that can perceive and act on screens.

GUI agents are a fast-growing niche in applied AI: rather than just describing an image, these models are trained to interpret app layouts, buttons, and text fields, then plan and execute the taps or clicks needed to complete a task. The vision component is what lets the model "see" an interface the way a user would, which is essential when there's no clean API to work against.

## Why it matters

- At 4B parameters, GELab-Zero-4B sits in a size range that can plausibly run closer to the edge, which matters for mobile control where latency and privacy are concerns.
- Building on Qwen3-VL rather than a proprietary base signals Microsoft's continued use of open weights as a starting point for specialized agents.
- The "preview" and "Zero" labeling suggests this is an early, experimental checkpoint rather than a finished product.

Microsoft has released the model under an "other" license, and key details such as context length and formal benchmarks aren't specified in the record. As a preview, it's best read as a research artifact and a signal of where Microsoft's GUI-agent work is heading rather than a production-ready release.

## Get the model

- [Hugging Face](https://huggingface.co/microsoft/GELab-Zero-4B-preview-Sico-Evolution)

## Sources

- [microsoft/GELab-Zero-4B-preview-Sico-Evolution](https://huggingface.co/microsoft/GELab-Zero-4B-preview-Sico-Evolution) — Hugging Face, Jun 30, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "Microsoft previews GELab-Zero-4B, a compact GUI agent", The Open Weights, Jun 30, 2026, https://theopenweights.com/news/gelab-zero-4b-f2ea