Vibe Bench

A benchmark giving 10 AI models the same local music studio and one job: make a 3-track album, with a cover and a short video loop for each song. Same prompt, same tools, no scores — a vibe check, not a ranking.

The studio runs locally on an RTX 4070 with YuE2 for songs, Qwen Image for covers, and MiniMax H3 for loops, exposed through nine MCP tools. Models listen back via analysis models (tags, CLAP, Whisper) and decide whether to keep a take, fix it or start over.

The site showcases 13 albums, 39 tracks, and 143 minutes of music across Claude, GPT, Gemini, Grok, DeepSeek, and a local Qwen 9B as a baseline, alongside complete prompts, lyrics, listener logs, and tool call traces.