Everything I publish starts on a single PC with an RTX 4080 and 12 GB of video memory. Ollama runs Qwen3 14B for the writing, ComfyUI runs Flux.1 for images, and a local text-to-speech model handles one of the narrator voices. None of it bills per use. Generating ten drafts costs the same as generating a thousand, apart from electricity.
A 14B model on 12 GB is good enough to draft and sometimes sloppy. It repeats itself, it likes certain phrases, and every so often it states something wrong with total confidence. That is why a stronger model sits at the end of the line as a reviewer. It reads what the local model wrote and says whether it can go out. The call is short, around 750 tokens, so the review adds almost nothing to the bill, and the heavy lifting stays free.
Local also means I can keep a pipeline running at 3 a.m. without watching a usage dashboard. When I'm testing a change to the video generator I can rerun it fifty times in an evening and the only cost is waiting. With a paid API I would have started rationing those runs by the third one, and I'd have shipped worse work.
The price is time and care. Flux images take a while to render, the models share one graphics card so jobs queue instead of running side by side, and updates to any of the tools can break something that worked last week. I keep the setup small on purpose: three main tools, each doing one job.
For a solo project that publishes every day, that trade has been worth it. I'd choose it again.
Top comments (0)