Kyutai's speech model reaches 77.1% on GSM8K

Kyutai’s Voice of Reason STITCH checkpoint scores 0.771 (77.1%) on GSM8K, reasoning while speaking instead of routing audio through a text-model cascade, according to the released model card. The score covers 1,310 written test items judged by GPT-4o, so it is a math benchmark, not proof of real-time conversational quality; Qwen Omni’s audio-agent budget makes the same distinction between cheap audio and cheap agents. Builders can file it as a credible open checkpoint to test, not a production-latency claim.