[New Model]: SongGeneration 2 / LeVo 2 (Tencent AI Lab) — Hybrid LLM-Diffusion Full-Song Generation

The model to consider.

SongGeneration 2 / LeVo 2 by Tencent AI Lab is an open-source full-song generation foundation model with hybrid LLM-Diffusion architecture. The v2-large checkpoint (4B params) was released 2026-03-01.

Model	Max length	Languages	GPU memory	RTF (H20)
SongGeneration-base	2m30s	zh	10G / 16G	0.67
SongGeneration-base-new	2m30s	zh, en	10G / 16G	0.67
SongGeneration-base-full	4m30s	zh, en	12G / 18G	0.69
SongGeneration-large	4m30s	zh, en	22G / 28G	0.82
SongGeneration-v2-large	4m30s	zh, en, es, ja, etc.	22G / 28G	0.82

License on the GitHub repo is NOASSERTION (custom Tencent terms in the repo / model cards — needs to be reviewed before integration). All RTF / memory numbers are quoted from the official README.

The closest model vllm-omni already supports.

There is no full-song / multi-minute music model in vllm-omni today. The closest existing pipelines are:

• Qwen3-TTS / Qwen3-Omni Talker for the AR stage (LeLM "composer brain" plays a similar role).
• Diffusion x2audio stages for the Hi-Fi renderer.

LeVo 2's architecture would map onto a multi-stage graph combining both.

What's your difficulty of supporting the model you want?

Architecture overview (from official README / project page)

• Hybrid LLM-Diffusion:
  • LeLM ("composer brain"): language model that manages global musical structure and performance details.
  • Diffusion ("Hi-Fi renderer"): synthesises high-fidelity acoustic detail conditioned on the LM output.
• Hierarchical Language Model: parallel modelling of
  • Mixed Tokens: high-level semantics (melody / structure).
  • Dual-Track Tokens: separate vocal and accompaniment streams for fine-grained acoustic detail.
• Training: 3-stage post-training — SFT on curated high-quality songs, large-scale offline DPO (~200k positive/negative pairs), then semi-online DPO driven by an automated aesthetic evaluator.

Reported results (per official README, requires independent verification)

• PER 8.55% on the team's evaluation (vs Suno v5 12.4%, Mureka v8 9.96% as cited).
• Subjective quality compared to commercial systems via 20-expert / 6-dimension / 100-songs-per-model evaluation.

Key integration challenges

• Multi-stage graph: AR LeLM stage + diffusion renderer stage + codec/VAE — vllm-omni already supports multi-stage graphs but this would be the first "music-scale" pipeline.
• Long-form output: up to 4m30s. Will exercise streaming + caching paths much harder than current TTS clips.
• External assets: ckpt/, third_party/, and per-version model checkpoints are downloaded separately. Some optional flags (--low_mem, --not_use_flash_attn, --separate / --vocal / --bgm) need to be surfaced in serving configs.
• License: needs review (NOASSERTION on the GitHub repo) before any redistributable build.
• Pending releases per the official TODOs: Music Codec / VAE, finetuning scripts, automated aesthetic evaluator, and v2-fast / v2-medium variants are not yet open.

Use case and motivation

1. Strongest open-source full-song model at time of writing, with reported PER and human evals against Suno / MiniMax / Mureka.
2. Multi-lingual lyrics: zh, en, es, ja and others in v2-large.
3. Controllable: text descriptions, style tags, prompt audio (10s reference), and structured lyrics with [Verse] / [Chorus] markers.
4. Active project with frequent releases (v2-large 2026-03; v2-medium / v2-fast and music codec/VAE on the roadmap).

References

• GitHub: https://github.com/tencent-ailab/SongGeneration
• Paper (v1): https://arxiv.org/abs/2506.07520
• HF org: https://huggingface.co/tencent/SongGeneration
• HF v2-large: https://huggingface.co/lglg666/SongGeneration-v2-large
• v2 demos: https://levo-demo.github.io/levo_v2_demo/
• HF Space playground: https://huggingface.co/spaces/tencent/SongGeneration
• Data pipeline (SongPrep): https://github.com/tencent-ailab/SongPrep

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.