Suggestion: Add MiDashengLM-0.6B for audio captioning

[RicherMans]

Hey! Love the smol-audio initiative — making audio AI models practical and accessible is exactly the right direction.

I noticed you have a notebook for fine-tuning Audio Flamingo 3 for audio captioning. While AF3 is a great model, it's quite large for on-device or lightweight deployment scenarios. I'd like to suggest adding MiDashengLM-0.6B as a compact alternative that fits perfectly with the "smol" philosophy.

Performance is more or less in line with some 7B models.

Benchmark	Qwen2.5-Omni-7B	MiDashengLM-7B	MiDashengLM-0.6B
MECAT-Caption (Long)	61.10	72.50	75.60
MECAT-Caption (Short)	56.50	72.30	74.70
MECAT-Caption (Mixed Sound)	23.80	23.00	42.40
MusiCaps	43.71	59.11	60.70
SongDescriber	45.31	46.62	51.90

Resources

• Model: https://huggingface.co/mispeech/midashenglm-0.6b-fp32
• Code: https://github.com/xiaomi-research/dasheng-lm
• Paper: MiDashengLM (arXiv 2508.03983)
• Training data: ACAVCaps
• Evaluation: MECAT benchmark

Would be great to see a notebook showcasing fine-tuning or inference with this model — it's a much better fit for the "smol" theme than the multi-billion parameter alternatives. Happy to help if needed!

If needed, I can PR an example .ipynb

Kind regards,
Heinrich