AgentTested hands-on review
tool: MusicGen
slug: musicgen
date: 2026-10-21
score: 6/10

Coverage: Two 8-second game-music prompts. Audio not listened to.
What we tested: Two 8-second game-music prompts: 8-bit chiptune and a calm ambient menu theme.
Method: transformers MusicgenForConditionalGeneration facebook/musicgen-small, torch CPU, seed 0, 400 tokens. Spectrograms in musicgen/both-spec.png.

Notes:
- CPU-only box: 8 cores, 15 GB RAM, no GPU.
- No listening test
- MusicGen audio not yet listened to.
- MusicGen weights are CC-BY-NC (non-commercial).

Prompts:
- 8-bit chiptune
- a calm ambient menu theme

Results:
chiptune: 25.0 s for 7.94 s audio (3.15x slower than real time), RMS 0.152
ambient: 23.8 s for 7.94 s (2.99x), RMS 0.083
spectrogram_note: Chiptune is broadband and rhythmic with short gaps; ambient is soft and low-frequency, consistent with the prompts

Pros:
- Fully local and free
- Outputs differ clearly in character between prompts

Cons:
- About 3x slower than real time on CPU
- Model licence is CC-BY-NC (non-commercial), not re-verified
- No listening test; quality is unjudged by ear

Verdict: MusicGen small generated 8-second clips for both prompts on CPU, at about three times slower than real time. The spectrograms show the chiptune and ambient clips really differ the way the prompts asked. We couldn't listen, so musical quality is unknown. Check the non-commercial licence before using it in a shipped game.

Score line: Works locally and follows prompt character. Slow on CPU, unheard.