DeepSeek V4 Pro 0813: coding tests with Hermes Agent

Fahd Mirza tests DeepSeek V4 Pro 0813 through Hermes Agent on two coding tasks. Although the video title says "Tested Locally," the walkthrough uses API access. It does not demonstrate how to run models locally, and the speaker says the weights will be released soon.

The first task involves a broken report-generation application with a FastAPI backend, a Redis pub/sub worker and a SQLite store. Submitted jobs remain pending. Mirza gives the model the repair goal without diagnostic hints, then checks the application in a browser. He reports a successful fix and demonstrates an inventory report progressing through processing to completion.

The second task asks for one self-contained HTML file using raw canvas and JavaScript, without libraries or a build setup. The requested scene is a lamb roast rotating above flames, with controls for cooking and seasoning. Mirza says generation took about 16 minutes. He demonstrates flame adjustment, spices and lemon, but says the spin movement could be better and that he expected more from the result.

The benchmark discussion cites Terminal Bench 2.1 rising from 72 to 88 and CyberGym from 53 to 83. Mirza explicitly identifies these as the provider's reported numbers; the two demonstrations are separate examples of the model's behavior as a coding assistant.