Feature: Add MiniMaxAI/MiniMax-Music3 support

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:

• MiniMax-Music3 model card
• Pinned model card
• Pinned pipeline metadata
• Pinned reproducible example
• Pinned license
• Upstream source repository

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.