Code-as-World-VL-4B
Code-as-World-VL-4B (https://arxiv.org/abs/2608.27549) is a vision-language model fine-tuned for physical understanding and quantitative reasoning over videos.
Project page: https://mirros-lab.github.io/code-as-world
GitHub: https://github.com/mirros-lab/code-as-world
Model details
- Base model: Qwen/Qwen3.5-4B
- Weight format: BF16 Safetensors checkpoint
- Recommended video input: 16 frames
Usage
The checkpoint can be served with vLLM:
pip install "vllm==0.19.1" "transformers==5.11.0" qwen-vl-utils
vllm serve MirroS-Lab/Code-as-World-VL-4B \
--served-model-name code-as-world-4b \
--max-model-len 4608 \
--gpu-memory-utilization 0.90 \
--media-io-kwargs '{"video":{"num_frames":16,"fps":-1,"video_backend":"openpangu"}}' \
--mm-processor-kwargs '{"do_sample_frames":false}' \
--mm-processor-cache-gb 0 \
--generation-config vllm
The server exposes an OpenAI-compatible API at /v1.
Intended use
This model is intended for research on physical understanding, measurement, and quantitative reasoning from images and videos. Model outputs may be inaccurate and should be independently verified before use in safety-critical settings.
License
This checkpoint is released under the Apache License 2.0. It is derived from Qwen/Qwen3.5-4B; users must also comply with the terms applicable to the base model and their input data.
Citation
@article{mirros2026codeasworld,
title = {Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning},
author = {{MirroS Team}},
journal = {arXiv preprint arXiv:2608.27549},
year = {2026}
}