The article spotlights LLM Reasoning, a feature that lets models perform step‑by‑step thinking in a single request. Developers tested it on local LLMs, notably gpt‑oss:20b running through Ollama on an AMD Ryzen 7 7700 with a GeForce RTX 4070. The focus is a maze‑solving experiment.

The maze test compared three prompt strategies: list, graph, and simple. The list format, a raw coordinate list, yielded the fastest and most accurate pathfinding, while the graph adjacency list worked only on small grids, and the human‑readable simple ASCII representation lagged behind. Accuracy hovered above 80 % for the list.

To inspect reasoning, the author shared Node.js and Python snippets that call the chat endpoint with a think parameter set to medium. The response includes a thinking field that shows the model’s internal chain, followed by the final content answer. The example demonstrates how the model articulates its deduction step by step.

The author concludes that Reasoning isn’t magic but disciplined word‑based logic, offering developers a ready‑made thinking engine. By observing the thinking content, teams can tweak prompt representations and even integrate function calling or tool use for more complex tasks. The code lives on GitHub for replication.