DeepSeek-GRM-27B is a reward model — it doesn't generate answers so much as it judges them, scoring outputs for quality and alignment. At 27 billion parameters, it sits in a practical middle ground: large enough to make nuanced distinctions between response quality, compact enough to run without extreme hardware. It's a specialist, not a generalist, built for the pipeline work of training and evaluating other models rather than direct user interaction.