WorldReward-qwen35-9b-GGUF

WorldReward-qwen35-9b is a 9-billion-parameter reward model built on Qwen3.5-9B, designed for reward modeling of camera-conditioned world models — given an input scene image, a text caption, a sequence of camera/movement actions (e.g., forward, left, camera_down), and a pair of candidate generated videos, it judges which video better satisfies the specified action trajectory, appearance quality, and motion quality, optionally producing explicit reasoning alongside its verdict. Evaluated on the WorldReward-Bench (760 human-labeled pairs, with strict three-way agreement scoring where "tie" predictions must match), it achieves the best reported agreement with human judgments across all three axes — 77.63% on Action, 81.32% on Appearance, and 73.03% on Motion — outperforming larger general-purpose models like GPT-5.5 and Gemini-3.1-Pro as well as specialized baselines like DAv3, WorldMirror, HPSv3, and UnifiedReward variants, and substantially exceeding zero-shot Qwen3.5-VL-27B and -9B baselines. The model requires a vLLM build that registers the Qwen3_5ForConditionalGeneration architecture and is used via the accompanying WorldReward GitHub repository's inference scripts, with the corresponding project page, model collection, and paper hosted separately; it is released under the Apache 2.0 license.

GitHub — https://github.com/CodeGoat24/WorldReward

Model Files

llama.cpp

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