OS-Helix-27B

OS-Helix-27B is a fine-tuned vision-language model based on Qwen/Qwen3.5-27B, released in Hugging Face Transformers format for screenshot understanding and computer-use agent research.

Property	Value
Base model	Qwen/Qwen3.5-27B
Architecture	Dense (27B)
Transformers class	Qwen3_5ForConditionalGeneration
Weight precision	BF16
Format	Sharded Safetensors
License	Apache 2.0

Quick start

Install PyTorch and torchvision for your CUDA environment, then install the inference dependencies:

pip install "transformers>=5.3.0" accelerate pillow

import torch
from PIL import Image
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "YYangzzzz/OS-Helix-27B"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)
messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": Image.open("screenshot.png").convert("RGB")},
        {"type": "text", "text": "Describe this screenshot and identify its main UI elements."},
    ],
}]
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=False,
    return_dict=True,
    return_tensors="pt",
).to(model.device)
with torch.inference_mode():
    generated = model.generate(**inputs, max_new_tokens=512)
answer = generated[:, inputs["input_ids"].shape[-1]:]
print(processor.batch_decode(answer, skip_special_tokens=True)[0])

The example uses non-thinking mode for a screenshot task. The packaged chat template supports the Qwen3.5 thinking switch. Choose the thinking setting appropriate to your application.

Release contents

The repository includes model weights and their index, model and generation configurations, tokenizer and chat template, image/video processor configurations, license and attribution, and SHA256SUMS for file integrity verification.

The selected fine-tuned checkpoint contains the main generation model without MTP weights. Use standard generation; MTP speculative decoding requires a separate compatible checkpoint.

sha256sum -c SHA256SUMS

The model configuration retains the upstream context setting; the fine-tuning sequence length is listed above. Longer-context behavior and downstream performance require application-specific evaluation.

Attribution

This release derives from the Qwen3.5 model family. See LICENSE for license details.

Citation

If you find this work helpful, please consider citing us:

@misc{yang2026oshelix,
  title = {OS-Helix: Unifying GUI Interaction and Code Execution in Native Computer-Using Agents},
  author = {Bowen Yang and Kaiming Jin and Zichen Ding and Zhenyu Wu and Zehao Li and Zhoumianze Liu and Kanzhi Cheng and Fangzhi Xu and Yanan Sun and Qiushi Sun and Yu Qiao},
  year = {2026},
  url = {https://arxiv.org/abs/2610.22234}
}