Kyutai releases Voice of Reason: speech-native models solve spoken math

Kyutai has released Voice of Reason, two open-weight speech-to-speech models built on GLM-4-Voice-9B that solve math problems out loud without a transcription step. Accuracy on spoken GSM8K rises from 27.3% to 77.1% via supervised fine-tuning and reinforcement learning; both checkpoints run on a single H100.

• Spoken GSM8K accuracy rises from 27.3% to 77.1% after SFT and RL
• glm-4-voice-of-reason-9b and stitch-9b are open on Hugging Face
• Training used 16 H100 GPUs, 150,616 Orca-Math problems and 1,500 RL updates
• Without temperature correction accuracy collapsed from 65.5% to 12.3%