Kyutai Releases Spoken-Math Models; Its Stronger Version Scores 77.1% in Its Test

Kyutai has released two open-weight Voice of Reason models that take spoken math questions and answer aloud without sending them through a separate text language model. In Kyutai’s test, the stronger version answered 77.1% of questions correctly, compared with 27.3% for the model it was built from.

How the model reasons while speaking

Both versions start from GLM-4-Voice-9B. Kyutai trained them first on examples of math problems, then with reinforcement learning that rewards answers judged correct. They handle speech directly when answering, rather than passing a transcript through a separate text model and converting its reply back to audio.

The direct version begins its answer without generating extra hidden reasoning. The STITCH version inserts silent reasoning chunks between spoken sections. While one section plays, it can prepare a later one. Kyutai says that overlap adds no extra delay, although the layout described for this release places a reasoning chunk before speech begins.

27.3%Starting model

Kyutai reports 27.3% for the original GLM-4-Voice on spoken GSM8K math questions.

70.3%Direct-answer release

The released version without extra hidden reasoning scored 70.3% in Kyutai’s test.

77.1%STITCH release

The released version with silent reasoning chunks scored 77.1% in Kyutai’s test.

Training accounts for much of the gain

The direct version’s score shows that reasoning during playback cannot explain the full improvement. For training, Kyutai used 150,616 problems drawn from Orca-Math. A text model rewrote the material for speech, and a speech system produced audio in multiple voices. A Qwen model then judged whether answers were correct during reinforcement learning; it did not see the reference answers.

The headline scores belong to the released checkpoints under a particular decoding setup. Across three training runs, the paper reported lower averages: 65.5% for the direct approach and 74.8% for STITCH. Kyutai also tested what happened when generated answers became audio. After that audio was transcribed, STITCH scored 72.0% following reinforcement learning, up from 66.2% after the earlier training stage.

A math specialist developers can host

The math gains did not carry over to every task. On spoken TriviaQA, a general-knowledge test, the starting model scored 40.6%; the trained direct version scored 34.0%, and STITCH scored 21.4%. The researchers attribute most of the decline to the first, full-data training stage. Larger alternative systems also scored higher on math in Kyutai’s comparison, though their size, design or output methods differed.

Both checkpoints are available as weights for self-hosting and retain the GLM-4-Voice license. Kyutai says each runs on one H100 GPU, but an audio application also needs components from the original GLM-4-Voice project to process incoming speech and turn generated audio codes into sound. The release offers developers a spoken-math model to run themselves, not a hosted voice service.