I made a video where I gave the exact same build job to DeepSeek V4 Flash twice, once on my own Mac and once in the cloud, and tracked every minute without cutting. Here is the gist and my take. If you want the full uncut session, the demo carries it better than words.
The job and the two runs
The app was small. You give it one video idea, it returns nine ranked titles, you pick one, and it generates a thumbnail. The twist in the prompt was that the app could not call any hosted LLM or image API. It had to drive Codex Exec to write the titles and make the images off my Codex subscription instead.
The local run took two hours and still needed some help from GPT to finish. The cloud run, using the Pi coding agent with DeepSeek V4 Flash through OpenRouter, had the same app working end to end in about 15 minutes. In the video you watch it happen in real time, including the server dying once mid-run and the model recovering on its own.
The honest asterisk
I want to be upfront about the quantization. The local copy ran at 2-bit and the cloud copy at a higher bit width, so the cloud version was both faster and working from a more accurate model. Not a fair fight, and I say so on camera. Speed came down to tokens per second, and the top OpenRouter providers for this model ran around 71 tokens per second in my run.
What it cost
The full build burned about 569,000 tokens at a 30 percent cache hit rate. At the roughly 10 cents per million input and 20 cents per million output I saw on OpenRouter, that came out under 15 cents for the whole task, crash and recovery included. Later I swapped the image step from Codex Exec to Gemini using keys already in my .env, and both title and thumbnail generation got noticeably faster.
My take is simple once you watch it run. "Open" does not have to mean "local." DeepSeek V4 Flash wants something like 96 to 128 gigabytes of unified memory to run at home, a three to five thousand dollar machine, or you can rent it by the token through OpenRouter for cents. The code and Pi config are on GitHub.
Work with me
I specialize in helping people and teams decide where their models should run and what that choice really costs. If you are weighing local hardware against a cloud bill, book a call at cloudyeti.io/meet and bring your actual workload.
Book a call