Code-as-World-VL-9B-GGUF

Code-as-World-VL-9B-GGUF

Code-as-World-VL-9B is a vision-language model from MirroS-Lab, fine-tuned from Qwen3.5-9B for physical understanding and quantitative reasoning over videos, as introduced in the accompanying paper (arXiv:2608.27549). Released as a BF16 safetensors checkpoint, it's designed to take video input at a recommended 16 frames and reason about measurement, physical properties, and quantitative relationships depicted in images and videos, rather than general open-domain chat, serving as the larger sibling to Code-as-World-VL-4B in the same model family. It's served through vLLM with an OpenAI-compatible API, using a fixed 16-frame, non-sampled video processing configuration and a modest 4,608-token max context length tuned for this task. Intended strictly for research on physical understanding and quantitative visual reasoning, its outputs may be inaccurate and should be independently verified before any safety-critical use, and it's released under the Apache License 2.0, inheriting the licensing terms of its Qwen3.5-9B base model.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp