Kyutai releases speech-native models for spoken mathematical reasoning
Kyutai has released two open-weight speech-to-speech models that solve spoken math problems without first transcribing the input or using a separate text language model. Both are based on GLM-4-Voice-9B and were trained with supervised fine-tuning followed by reinforcement learning. On spoken GSM8K, the base model’s accuracy increased from 27.3% to 77.1% for the released Stitch checkpoint.
The first model, glm-4-voice-of-reason-9b, answers directly while speaking its calculations. The second, glm-4-voice-of-reason-stitch-9b, inserts silent 100-token reasoning segments between spoken passages; Kyutai says later segments can be generated while earlier audio is playing. The released checkpoints scored 70.3% and 77.1%, respectively, without the paper’s top-k decoding setup.
Supervised training used 150,616 speech-adapted problems from Orca-Math and raised accuracy to 61.7%. Reinforcement learning used four sampled responses per question and a binary reward from Qwen3-235B-A22B-2507. Training involved 16 H100 GPUs and 1,500 updates. The models can be self-hosted on a single H100, but users also need the GLM-4-Voice repository’s speech tokenizer and decoder. Their weights use the GLM-4-Voice license.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.