Vision-language modeling with recurrent visual computation
This repository contains **one `model.safetensors` and flat configuration, image-processor and tokenizer files**. All inference and evaluation code is kept in the [LoopVL GitHub repository](https://github.com/Tier-Flow/LoopVL). There are no Python files in this model snapshot. ## Quick start Use Linux, Python 3.12 and a matching CUDA-compatible PyTorch/torchvision pair. The validated GPU environment uses torch 2.12.1 and torchvision 0.27.1. ```bash git clone https://github.com/Tier-Flow/LoopVL.git cd LoopVL python -m pip install -r requirements.txt hf download TierFlow/LoopVL --local-dir model # Alternative: ms-hub download Eternity123/LoopVL --local-dir model python scripts/verify_repo.py --verify-model python runtime/infer.py --image /path/to/image.png --prompt "What is in the image?" --budget 32 --device cuda:0 ``` Keep **both the GitHub code and all downloaded model files**. LoopVL uses its custom GitHub loader. The main checkpoint uses H2L3: `L → L → L → H → L → L → L → H`, with 16 layers per module call and **128 effective layer applications per forward pass**. Preserve its mixed BF16 core weights and FP32 adapters; do not cast the entire model with `.half()` or `.bfloat16()`. > [!TIP] > For LoopVL-1B benchmark evaluation and behavior exploration, prefer direct > answers. Chain-of-thought prompting is recommended only for mathematics tasks.