DeepScaleR is a newly released model that is designed to exceed the performance of the O1-Preview model, utilizing reinforcement learning to achieve this goal with 1.5 billion parameters. There is a division in opinions about the efficacy of large models vs. specialized smaller models. Some commenters caution that reliance on current benchmarks can be misleading and advocate for the evaluation of models based on user-specific KPIs. End-users are experimenting with DeepScaleR for tasks ranging from reasoning prompts to model evaluations, noting both successes and limitations in its performance. The overall sentiment is a call for more robust training methods and trust in benchmark assessments, alongside a desire for models that can handle more complex tasks effectively.