Step Audio R1.1

Step Audio R1.1 is an audio reasoning checkpoint for spoken interaction, with a formulation and articulation architecture.

Reasoning grounded in speech

Step Audio R1.1 separates a formulation component for reasoning from an articulation component for spoken delivery. The publisher describes thinking while speaking and reasoning over acoustic representations. The downloadable repository’s main inference path uses audio inputs and returned text, while the companion interaction system supplies spoken delivery.

Check the complete interaction stack

Use the supplied serving and client instructions when testing realtime behaviour. A text-generating audio inference route does not alone establish a complete bidirectional speech system. Match the checkpoint identifier, audio encoding and chat template to the selected path.

Listen and verify meaning

Try a spoken request with a correction, a question whose tone matters and a task that includes background sound. Check whether the response follows the final intent and distinguishes the audio evidence from assumptions. For the companion voice workflow, measure the first audible response and interruption handling separately from the text generation speed.

Self-hosting scope

The released configuration sets 65,536 positions and the full checkpoint contains roughly 33.5B parameters. The weights use Apache 2.0 licensing. This profile describes the downloadable model and keeps public API token rates unknown. A StepFun studio or another managed speech route may expose additional components and different limits, which should be verified through its own documentation.

Sources

Model documentation