HappyShrimp targets full-song coherence with hierarchical planning and flow matching
The new song product arrives beside an Alibaba Token Foundry technical report built around global planning, local token generation and full-song audio rendering. The public paper does not itself name HappyShrimp.

HappyShrimp has opened a song-generation product built around prompts, custom lyrics, instrumentals and optional reference audio. Its public site combines generation with a social discovery feed where examples and prompts can be shared. The product footer also names Taihe Music Group as a strategic partner.
The launch is presented alongside a technical system from Alibaba Token Foundry designed for full-song coherence. Its research paper describes an 8-billion-parameter global language model that plans long-range structure, a 0.4-billion-parameter local model that completes music tokens, and an 8-billion-parameter FullDiT renderer that uses flow matching with access to the full song context. The reported output is stereo audio at 48 kHz.
One identification boundary is important: the paper calls the evaluated system Lucky Dolphin, its anonymous entry on Artificial Analysis, but does not use the product name HappyShrimp. The authors report an Elo score of 1,129 and a rank interval of second to third with overlapping confidence bounds. Those are the paper's evaluation results, not independent proof that the public product is the same checkpoint or will produce the same results.
Alibaba now has several music-model lines
The research sits beside two other public efforts. The Qwen-Music paper describes a separate music-generation approach using a single-codebook tokenizer and Melody-CoT training. Alibaba Cloud's Fun-Music service exposes Chinese- and English-language song and instrumental generation through Model Studio, with streaming and non-streaming API modes. Fun-Music is currently documented as a limited preview in the China (Beijing) region.
HappyShrimp's product claim is that long songs should retain melody, arrangement and development instead of optimizing only a convincing short passage. The architecture in the Token Foundry paper is designed for that problem. A defensible product conclusion will require listening tests across complete songs — and a direct, primary-source statement that maps the public HappyShrimp release to the reported Lucky Dolphin system.
Related
| AI music tool | Fun-Music |
|---|---|
| AI music tool | HappyShrimp |
| AI music model | HappyShrimp 1.0 |
| Company or organization | Taihe Music Group |
| Company or organization | Alibaba Token Foundry |
| Company or organization | Alibaba Cloud |
| Research | Qwen-Music Technical Report |
| Research | Pushing the Frontier of Full-Song Generation: Hierarchical Autoregressive Planning Meets Flow-Matching Rendering |



