A team of interns from the University of Waterloo has released a new open-source artificial intelligence model after distilling approximately 2.3 million reasoning traces generated by Anthropic's Claude Fable 5. The project demonstrates how knowledge from a much larger AI system can be transferred into a smaller, more efficient model while preserving highly consistent reasoning behavior.

Built on the Qwen3-4B architecture, the model contains around four billion parameters and was trained using distilled reasoning data collected from Claude Fable 5. Rather than recreating the original model, the project focused on teaching a compact open-source model to reproduce similar reasoning patterns across a wide range of tasks.

During testing, the researchers reported unusually high levels of consistency. According to the published benchmark results, the model achieved 100 percent self-consistency when evaluated with 512 sampled responses, producing identical answers across repeated runs. The team also reported zero output entropy and no measurable hallucination variance during its experiments, suggesting the model consistently generated the same reasoning path under identical testing conditions.

One of the project's most discussed findings was the model's repeated tendency to converge on a single response across evaluations. In demonstrations released by the researchers, the model consistently returned the same conclusion, illustrating what the team described as an extreme level of reasoning stability after large-scale distillation.

The release has attracted attention from the open-source AI community, where researchers are closely watching new approaches that improve the performance of smaller language models without requiring massive computing resources. Knowledge distillation has become an increasingly popular technique because it enables lightweight models to inherit capabilities from significantly larger systems while reducing deployment costs.

The Waterloo team's decision to make the model publicly available is expected to encourage further research into AI reasoning, consistency and model compression. Developers can study the training approach, reproduce experiments and evaluate how distilled reasoning affects model behavior across different applications.

The project also contributes to the growing trend of open-source AI development, where academic researchers and independent developers are publishing increasingly capable models alongside their methods and evaluation results. As interest in efficient AI systems continues to grow, experiments like this could help shape future techniques for building smaller models that deliver reliable and repeatable performance without requiring frontier-scale infrastructure.