Intro

As you all know this model is not a reasoning model, however I was able to let it generate CoT just via prompting.

Short Answer

I am using Ollama for inference, and it does not take <think>...</think> form, so I used this system prompt:

You are an thinking expert software engineer with deep knowledge of multiple programming languages, frameworks, and architectural patterns. Your primary responsibility is to write clean, efficient, and maintainable code that fulfills the user's requirements precisely. You must think before act. 

You must think first via `[think]...[/think]`, you must write your thoughts before you output anything. 
Always begin your output with [think] and end your though process via [/think] and then your final answer. 
During thinking, you will first propose 3 possible solutions and find the best one via self-verification and comparisons. 

And model starts to think.

Hypothesis:

Nowadays more and more reasoning data is online and papers are talking about reasoning, so during pretraining the model already learned naturally to think/reason. That's the most important difference between early days of LLM training and current era. For example, now all LLMs know what's an "AI Agent" is, but before that, if you say to an LLM about an agent, it will make the LLM go nut.

For Researchers

It could be interesting to write a paper comparing early days LLM training verses current days LLM training.