Frontend · Backend · Artificial Intelligence

MiniMax Music-3 Runs on a 16 GB RTX 5060 Ti: Local AI Music That Actually Sounds Like a Song

By 雪隐_上班了 · Aug 16, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Local music generation that doesn't sound plastic has been a missing piece for hobbyist AI rigs. A 16 GB card can now produce usable, license-clean vocal tracks in under two minutes, which lowers the barrier for game jams, podcast intros, and indie content creation without recurring API costs.

Summary

MiniMax Music-3, packaged for ComfyUI by Comfy-Org, produces full vocal tracks with lo-fi warmth, vinyl crackle, and natural song structure from text prompts. The full model repo weighs 61.8 GB, but three int8-quantized files totaling 11.1 GB fit comfortably within a 16 GB card's budget. Skipping the fp16 DiT in favor of the int8 version and enabling tiled VAE decoding are the two changes that prevent an out-of-memory crash mid-generation. The model interprets structural tags like [verse] and [chorus] as real compositional instructions rather than decorative labels, and it outputs MP3s around 4–5 MB under an Apache 2.0 license that permits commercial use of generated audio. GPU temperatures stayed at 72°C, a full 6–8°C cooler than video generation workloads on the same hardware.

Takeaways

— Only three int8-quantized files (11.1 GB total) are needed to run Music-3 on a 16 GB GPU; the full 61.8 GB repo is unnecessary.

— Swapping the DiT model from fp16 to int8 and enabling tiled_decode are mandatory to avoid CUDA out-of-memory errors on 16 GB cards.

— Structural tags like [intro], [verse], [chorus], and [outro] function as real compositional instructions, not decorative labels.

— The model outputs roughly 59 seconds when max_duration is set to 60, because it finds a natural latent-chunk boundary to end on.

— GPU temperature settled at 72°C, significantly cooler than the 78–80°C seen during video generation on the same hardware.

— Generated audio carries an Apache 2.0 license and permits commercial use, but cannot be used to train other AI models or impersonate real artists.

Conclusions

Music-3 treats structural tags as executable composition commands, which is a fundamentally different paradigm from Suno's verbatim lyric-to-vocal approach and gives more control over song architecture.

The 6–8°C temperature drop versus video generation suggests DiT-based audio models impose a lighter compute burden than video diffusion, making music generation a more sustainable local workload for mid-range GPUs.

Comfy-Org's decision to ship int8-converted weights alongside fp32/bf16 variants acknowledges the real-world install base of 12–16 GB cards, a pragmatic move that other model packagers often skip.

Concepts & terms

int8 quantization

A model compression technique that stores weights as 8-bit integers instead of 16- or 32-bit floats, roughly halving VRAM usage at a small quality cost. The 'convrot' variant applies an additional rotation to convolution weights to better preserve precision.

tiled VAE decode

A method that splits the latent representation into smaller spatial chunks before decoding, processing each tile sequentially to keep peak memory low. Essential for long audio or high-resolution images on VRAM-constrained GPUs.

DiT (Diffusion Transformer)

A diffusion model architecture that replaces the traditional U-Net backbone with a transformer, used here as the core generative component that iteratively denoises latent audio representations.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗