Skywork releases Reward-V2, eight reward models from 0.6B to 8B

For context: this is a mid-2025 release, outside the last twelve months, included because reward models are a core kind of decision model. The card says the paper was published on 2 July 2025 and the collection was last updated on 4 July 2025.

Skywork-Reward-V2 is a family of eight reward models for scoring conversation responses. The card says it uses the Bradley-Terry approach. The family is Llama-3.1-8B, an experimental Llama-3.1-8B-40M for research use, Llama-3.2-1B and 3B, and Qwen3 at 0.6B, 1.7B, 4B and 8B. The Llama-based models fall under the Llama 3.1 community licence. Training data is about 26 million preference pairs curated through a human-LLM pipeline, and the 40M variant uses 40 million pairs. Maximum input is 16,384 tokens.

For the 8B Llama model the card reports RewardBench v1 96.4, RewardBench v2 84.1, PPE Preference 77.3, PPE Correctness 83.4, RMB 86.4, RM-Bench 92.8 and JudgeBench 80.0, an average of 85.8. The maker is Skywork, part of Kunlun Inc.

A reward model returns a score, not an answer, which is why it counts as a decision model. These benchmark numbers are from the maker's card and paper.

Sources

• https://huggingface.co/Skywork/Skywork-Reward-V2-Llama-3.1-8B
• https://arxiv.org/abs/2507.01352

Editor's first pass (AI-assisted)

• Real 4.5: Public weights for all eight models; Llama variants fall under the Llama licence.
• Useful 4: Scores responses for RLHF and best-of-N selection.
• Notable 3.5: A large open reward-model family with published benchmark scores.
• Novel 3: Scales preference data and curation rather than a new method.