Qwen 3.8 with Ollama: local coding build and comparison

Qwen 3.8 27B builds a YouTube title and thumbnail generator in this coding demonstration. The speaker runs the model through Ollama in VS Code with the Pi coding agent and reports roughly 27 tokens per second. The prompt asks for nine scored titles per video idea and a thumbnail for a selected title, using Codex Exec rather than direct model or image APIs.

The session follows environment checks, a Codex capability test, server code and frontend files. The model also adds an SVG thumbnail fallback, a health endpoint and a README without explicit requests. The speaker steers it about unnecessary work, so the claimed one-shot result includes some guidance, although he says he never fixes code manually.

The speaker reports a working app after about 26 minutes, compared with about 50 minutes for Qwen 3.6 27B and two hours for DeepSeek V4 Flash on the same prompt. These are his session results, not a general benchmark. A further prompt improves the interface, but he still finds rough edges and concludes it falls short of Opus 4.6.

The local LLM handles coding; the app delegates title and image generation to the speaker's Codex subscription. The claim of offline operation concerns the downloaded coding model, not an established end-to-end property of the app. Hardware details conflict: the title names a 128 GB M3 Mac, while the description names a 24 GB MacBook and the speaker cites 24 GB of VRAM.