I’m genuinely fascinated (and a little emotional) right now.

For over two weeks, I pushed a tiny Qwen 3.5 0.8B model locally to its absolute limits. I spent hours crafting better prompts, optimizing my code to send lighter requests, and iterating relentlessly.

Then I decided to try a cloud model: Grok 4.1 Fast.

The difference was shocking.

The quality jumped so dramatically that, for a moment, I felt like I had wasted my time with the small model. But that couldn’t be further from the truth.

Those two weeks of struggle with the 0.8B model taught me more about prompting, LLM behavior, and system design than I could have learned any other way. I tested everything: broad prompts vs hyper-specific ones, single complex prompts vs breaking tasks into tiny atomic steps, and I gained a deep understanding of what small models can (and cannot) do compared to larger ones.

The results with Grok are undeniably better… but the real value was in the grind.

Once again, it’s the path — not just the destination — that makes all the difference.