The way the big guys are feigning improvement right now is through recursion. They feed the LLM back into itself a few thousand/million times to narrow the normal curve of result accuracy. This is incredibly expensive, unsustainable, and not possible to do (in reasonable time frames) without data centers on data centers. Don’t expect open models to achieve the same kind of results with coding related queries because the underlying technology has plateaued.