Stepfun-Ai/Gelab Zero 4b Preview

On December 19, 2025, Hugging Face staff member nielsr merged a pull request improving the GELab-Zero-4B-preview model card, adding the image-text-to-text pipeline tag, the transformers library name, and prominent links to the Step-GUI Technical Report, the opengelab project page, and the stepfun-ai/gelab-zero code repository. The card's config.json indicates Qwen3VLForConditionalGeneration compatibility. This metadata-only change did not alter the model itself but formalized its discoverability on the Hub and surfaced the official research and engineering resources for the Stepfun-Ai GUI agent. It remains the most recent official activity documented for GELab-Zero-4B-preview, confirming its classification as a vision-language Android control model with transformers-compatible inference.

Stepfun-Ai's GELab-Zero-4B-preview is a 4-billion-parameter vision-language GUI agent released as part of the GELab-Zero project, packaged with plug-and-play inference infrastructure for ADB connections, dependency installation, and task recording/replay. The model is fine-tuned from Qwen3-VL-4B-Instruct and is distributed in BF16 safetensors shards (~4.99 GB plus ~3.89 GB), with a chat template and video preprocessor included. Capabilities highlighted in the model card include local deployment on consumer hardware, GUI navigation via click, type, slide, and wait actions, complex multi-step task execution across apps such as food delivery, transportation, shopping, and social media, and zero-shot generalization to unseen applications. An Ollama quick-start is documented for local inference, while the accompanying Step-GUI Technical Report, the opengelab.github.io project page, and the stepfun-ai/gelab-zero GitHub repository provide the full reference materials.

The LLM Pricing directory lists GELab-Zero-4B-preview as GA via the Qiniu gateway with an 8,192-token context window and a 4,096-token output limit. Documented capabilities include reasoning, tool use, structured output, temperature control, and image attachments, with multimodal text and image modalities. Released and Pricing is explicitly undisclosed with no paid channels listed, meaning developers cannot estimate cost from this aggregator and must rely on Qiniu's gateway compute pricing. The model is positioned as a StepFun flash model for efficient multimodal reasoning, coding, and tool use, and the aggregator is the sole seconda

OpenModelStats independently corroborates the model's core metadata: 4,437,815,808 parameters (~4.4B), Image-Text-to-Text task, Apache-2.0 license, transformers/safetensors stack with Qwen3-VL architecture, and derivation from Qwen/Qwen3-VL-4B-Instruct. Published November 28, 2025; last source update December 19, 2025. Adoption signals are modest and cooling: ~19.7K cumulative downloads, ~269 rolling 30-day downloads, ~17/day 7-day velocity, 154 likes, ranked #3489 among tracked Image-Text-to-Text models (down 272 places week-over-week). The declining rank suggests competitive pressure from other 4B multimodal models like Qwen3-VL-4B

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Run stepfun-ai/GELab-Zero-4B-preview with fast, reliable, and scalable inference on FriendliAI. Get low-latency performance with advanced quantization (FP4, FP8, INT4, INT8), continuous batching, optimized GPU kernels, token caching, and seamless API integration.

A 4B GUI Agent for autonomous Android device control with zero-shot generalization.

We’re on a journey to advance and democratize artificial intelligence through open source and open science.