Skip to content

Feature: Add MiniMaxAI/MiniMax-Music3 support #1092

Description

@yifeif-nv

Request type

New model or checkpoint support

Problem and use case

TensorRT-Model-Connect users cannot currently build and run MiniMaxAI/MiniMax-Music3 as an end-to-end text-to-music pipeline. The requested use case is to generate complete songs from lyrics plus a detailed music description while keeping the model execution and artifact flow inside TensorRT-Model-Connect.

Existing dense Qwen3 support covers only one part of this architecture. Music3 also requires its local RVQ decoder, condition encoder, flow-matching transformer, scheduler, Flow-VAE/vocoder path, and audio-specific input/output contracts.

Proposed outcome

Add qualified MiniMax-Music3 support with these acceptance criteria:

  • Recognize the exact public checkpoint and immutable revision below, and fail closed for incompatible layouts or revisions.
  • Build the required components at BF16 where supported, preserving higher precision where audio quality or upstream component contracts require it.
  • Package all required engines and metadata into a TensorRT-Model-Connect bundle.
  • Run end to end through the native TensorRT-Model-Connect runtime, without an upstream Python inference fallback.
  • Accept lyrics and a music description plus reproducibility controls such as seed and duration or maximum frames.
  • Produce a valid 32 kHz, 16-bit stereo WAV for the pinned upstream example.
  • Validate fixed prompts and seeds against the pinned upstream reference using model-owned, meaningful checks for audio validity, lyric intelligibility, musical structure, and perceptual quality. Do not weaken validation thresholds to make the test pass.
  • Document supported duration, prompt length, precision, memory, streaming, and GPU boundaries.
  • Record the legal/compliance disposition for the custom MiniMax-Music3 Community License before model onboarding or redistribution.

Model and target details

  • Model ID: MiniMaxAI/MiniMax-Music3
  • Revision: fbdf52fbaaca799592917417eb05f1899f1255ec
  • Task: text-to-music / text-to-audio from lyrics and a music description
  • Requested precision: BF16, matching the upstream Diffusers example, with component-specific precision exceptions where required
  • Initial qualification target: NVIDIA H100 80GB
  • Expected runtime path: standard TensorRT backend, bundled build, native TensorRT-Model-Connect execution
  • Upstream architecture: MiniMaxMusic3ForConditionalGeneration / model_type=minimax_music3
  • Major components: Qwen3-based global language model, local RVQ depth decoder, condition encoder, 1D flow-matching transformer, scheduler, and vocoder

Alternatives considered

  • SGLang-Omni exposes the model through /v1/audio/speech.
  • The upstream Diffusers modular pipeline supports local generation.
  • ComfyUI provides an integration.
  • Existing TensorRT-Model-Connect Bark, Magpie TTS, and dense Qwen3 paths do not provide this checkpoint's complete text-to-music pipeline.

These are functional workarounds, but they do not provide a qualified TensorRT-Model-Connect build, bundle, and native runtime path.

Additional context

Public references:

The model card describes an 8B global LLM, a 0.6B local LLM, flow-matching synthesis, up-to-five-minute generation, and 32 kHz stereo WAV output. Current upstream limits include non-streaming generation, a 5,000-token text-prompt limit, and a 9,000-frame audio-generation limit.

Submission checks

  • I searched open and closed issues and found no duplicate request.
  • I removed secrets, private/internal evidence, personal paths, and restricted artifacts.

Metadata

Metadata

Assignees

Labels

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions