Two speech-to-speech models arrived from Kyutai under the name Voice of Reason. Both are open-weight, and both do arithmetic directly in audio. No transcriber sits in the loop.
Each checkpoint descends from GLM-4-Voice-9B, then picks up fine-tuning and reinforcement learning on top. On the spoken GSM8K test, the base model started at 27.3 percent. The finished versions reach 77.1 percent. An earlier method called Stitch had managed 58.7 percent with reasoning chunks, so the team calls this the first dose of reinforcement learning applied to math reasoning in a speech-native model.
Scoring centres rewards within each group, giving an objective close to GRPO that drops PPO clipping and KL regularisation. Updates totalled 1,500. Sixteen H100 GPUs carried the run. Serving one checkpoint fits on a single H100 in BF16, and the weights inherit the GLM-4-Voice licence.
The trade-off is built in. Cascaded systems that chain a transcriber, a text model and a voice synthesiser still reason better, but every stage adds delay and strips out tone. Speech-native models must emit audio at steady intervals to feel conversational, which caps how much silent thinking they can afford.
The paper’s numbers use top-k 50 decoding averaged over three seeds, with GPT-4o acting as judge.
