VTR-2B

Visual Token Routing (VTR) on Qwen3.5-2B. A small router, trained jointly with the model, decides which visual tokens enter the language model. VTR matches the full-resolution model on OSWorld, CUA-World and video game action prediction with 2.5 to 5.8× fewer visual tokens.

• Paper: arXiv:2611.06854
• Base model: Qwen/Qwen3.5-2B
• Router: three-layer MLP with 23.3M parameters, about 1% of the language model
• Training: 2,000 steps of joint training
• Every 10th image is kept in full

Usage

The router runs through the patched vLLM server in the VTR repository.

git clone https://github.com/pranjal2041/vtr
cd vtr/inference
uv sync
uv run vtr patch
uv run vtr serve l3lab/VTR-2B --port 8000

The server is OpenAI-compatible. See inference/README.md for the fast mode and other options.