Falcon-H1R is a new 7-billion-parameter model designed to enhance reasoning capabilities in small language models (SLMs). It demonstrates that smaller models can achieve competitive reasoning performance compared to significantly larger state-of-the-art models, which are two to seven times bigger. The model excels in various reasoning benchmarks, thanks to a focus on data curation and strategic training methods, including efficient supervised fine-tuning and reinforcement learning scaling. This approach allows Falcon-H1R to deliver strong performance without the typical increase in model size. The architecture of Falcon-H1R incorporates a hybrid-parallel design, which contributes to faster inference and improved token efficiency while maintaining high accuracy. This combination makes it a valuable option for tasks that require extensive reasoning, particularly those that involve generating detailed chains of thought and parallel test-time scaling. The model also employs DeepConf, a new technique that enhances test-time scaling efficiency, yielding better accuracy and reduced computational costs. Falcon-H1R's design choices and training strategies highlight how compact models can still achieve robust reasoning performance in demanding scenarios.