I dunno. Maybe you're expecting too much of them? They're obviously not gonna be like those massive data-center LLMs, but I've had some pretty good "brainstorming" sessions about code and documentation with Qwen and Gemma, and the latest vision-capable Qwen models do a really decent job of extracting data and text out of images (often more accurately than typical OCR engines) and even describing or captioning images, and for writing code, I've had good success with asking it to write me docstrings and type hints, and common "boilerplate" code, and even some pretty solid function requests that often come out exactly as I expected it to (as long as it's fed good "context" to work with, like some basic code-style rules, and access to library documentation and existing codebase).