A 27M parameter model beating LLMs on reasoning tasks | Dark Hacker News