Step-Audio R1
Step-Audio R1 uses Modality-Grounded Reasoning Distillation to keep reasoning tied to acoustic evidence. The training method addresses a failure in which a model reasons about a transcript instead of the sound.
Restricted access
Use the model's serving path
The release provides a customized vLLM deployment path. The examples are specific to its audio inputs and prompt template.StepFun Step-Audio-R1 model card
Preserve the audio evidence
Practical guidance. Use recordings that require sound-level interpretation when evaluating this model. A benchmark based only on transcript text would not test the distinction that motivated its training.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Restricted access |
| Access | Downloadable |
| Model identifiers | Step-Audio-R1 |
| Access gate | Hugging Face requires acceptance of the repository's conditions and sharing of account information before file access.StepFun Step-Audio-R1 model card |
| License | The published card labels the model Apache 2.0.StepFun Step-Audio-R1 model card |
| Domains | The card covers speech, environmental sound and music reasoning.StepFun Step-Audio-R1 model card |
Related models
- Step-Audio R1.1
StepFun · Step-Audio R
An audio model that separates higher-level reasoning from speech generation through a dual-brain design.
Launch date not confirmedDownloadable - Step-Audio R1.5
StepFun · Step-Audio R
Step-Audio R1.5 adds human-feedback training for natural dialogue while reasoning and speech proceed together.
Launch date not confirmedAPI · Provider product