Kyutai Teaches Speech Models to Do Math Out Loud

🔥 What happened Kyutai released Voice of Reason, two open-weight speech-to-speech models that solve math problems out loud. No transcription step, no separate text LLM. On spoken GSM8K, accuracy jumps from 27.3% to 77.1%. 💡 Why it matters The recipe: supervised fine-tuning plus reinforcement learning on 16 H100 GPUs. One detail is critical – without temperature correction, accuracy collapses to 12.3%. Both 9B checkpoints run on a single H100, making self-hosting realistic. ⚡ Our take Speech-native reasoning without a text detour is a genuine paradigm shift. Anyone still betting on cascaded pipelines is sacrificing latency, paralinguistic cues, and soon, relevance.