GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
ayushadarsh7
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
ayushadarsh7
Model Tree
Qwen/Qwen3-VL-4B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Use Python 3.10+ and a compatible PyTorch/CUDA installation. requirements.txt records the versions used for the export validation.
bash
pip install torch torchvision transformers accelerate pillow safetensors huggingface_hubhf download ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B infer.py --local-dir kayak-inferencepython kayak-inference/infer.py --model ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B --image screenshot.png --instruction "Click the Search button."For a fill action, add --value "Chennai" and provide an instruction such as Enter Chennai as the departure city.
Download all model files for offline use:
bash
hf download ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B --local-dir kayak-modelpython kayak-model/infer.py --model ./kayak-model --image screenshot.png --instruction "Click the Search button."The inference script automatically uses a CUDA GPU when available. CPU inference uses more memory and is slower. A GPU with at least 16 GB VRAM is a practical starting point for a single screenshot; memory needs depend on image resolution and input length.
Screenshot grounding model fine-tuned on Kayak flight-search interfaces. Given a clean screenshot and a single-step instruction, it predicts JSON containing an action and a normalized bounding box, plus a value when applicable.
This is not an autonomous booking agent. It does not plan or execute browser actions.
bbox is [x, y, width, height] on a 0–1000 scale, with x/y at the top-left corner. For an image of width W and height H: pixel x = x/1000W, pixel y = y/1000H, and similarly for width/height. For clicking, use the box centre. Fill actions may also include value.
These are single-step offline grounding metrics, not end-to-end browser task success rates. They describe the original adapter evaluation, not a new full benchmark of the merged export.
Merge validation: standalone BF16 reload succeeded; 3/3 checked predictions exactly matched the unmerged adapter. See merge_validation.json. BF16 merging can introduce small numerical differences.
Fine-tuned on a small domain-specific dataset; UI changes, different screen sizes, and unfamiliar sites can reduce accuracy. Inspect generated coordinates before executing consequential actions. No booking, payment, or personal-data workflow was evaluated here.
This derivative uses the base model's Apache-2.0 license. Original base revision: ebb281ec70b05090aa6165b016eac8ec08e71b17. Original adapter SHA-256: adf546ef4902717cfbeca735e9f764688dbf3cb335279935ddfb448b70ae26e3. The upstream license is included when supplied by the base repository.
pip install torch torchvision transformers accelerate pillow safetensors huggingface_hubhf download ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B infer.py --local-dir kayak-inferencepython kayak-inference/infer.py --model ayushadarsh7/RuntimeLabs-Kayak-Qwen3-VL-4B --image screenshot.png --instruction "Click the Search button."