Qwen3-8B · MiMo Music GRPO — GGUF
Code, logs and all results: https://github.com/PromptEngineer48/mimo-music-grpo
GGUF builds of Prompt48/Qwen3-8B-MiMo-Music-GRPO (Qwen3-8B trained with GRPO on Xiaomi's open-sourced MiMo-V2.6 music environment). See that page for results and limitations.
Use Q8_0. Measured on the same 40 held-out prompts × 4 samples, Xiaomi's strict grader:
| Version | Size | Valid music | Mean score |
|---|---|---|---|
| Base Qwen3-8B (before RL) | — | 0.6% | 0.002 |
| Merged bf16 (vLLM) | 16 GB | 14.4% | 0.050 |
qwen3-8b-mimo-music-grpo-Q8_0.gguf |
8.7 GB | 18.8% | 0.071 |
qwen3-8b-mimo-music-grpo-Q4_K_M.gguf |
5.0 GB | 5.0% | 0.015 |
Q4_K_M loses most of the RL gain (the learned behaviour is fragile at 4-bit). It still fits 8 GB GPUs, but expect far fewer valid pieces.
Prompting matters: it was trained on long, detailed composition briefs (key, tempo, meter, instrumentation, bars, structure),
like the ones in XiaomiMiMo/MiMo-V2.6-RL-oss (music subset). Short prompts give short, repetitive output.
Ollama
ollama run hf.co/Prompt48/Qwen3-8B-MiMo-Music-GRPO-GGUF:Q8_0
Add /no_think to your message (trained with thinking off) and use temperature 1.0. System prompt used in training:
You are a composer. Reply with one complete piece in ABC notation inside a ```abc code block. Start with X:1 and include T:, M:, L:, Q:, K: headers.