Kyutai Releases Voice of Reason: A Speech-Native Model that Solves Spoken Math with Reinforcement Learning

Kyutai has released Voice of Reason, a set of 2 open-weight speech-to-speech models built on GLM-4-Voice-9B.

According to the release, supervised fine-tuning and reinforcement learning lift spoken GSM8K accuracy from 27.3% to 77.1%.

Key Highlights

• 2 open-weight speech-to-speech models built on GLM-4-Voice-9B.
• Spoken GSM8K accuracy improved from 27.3% to 77.1% through supervised fine-tuning and reinforcement learning.
• There is no transcription step and no text LLM in the loop.
• Both checkpoints are available on Hugging Face and run on a single H100.