Kyutai releases speech-native math models that skip the text LLM
The released GLM-4-Voice derivatives score up to 77.1% on GSM8K, though larger cascaded systems still lead the benchmark.
By Ryan Merket · Published
Primary source: Kyutai on X
Why it matters
Voice assistants usually trade responsiveness for reasoning quality by sending transcripts through a text model. Kyutai shows that reinforcement learning and overlapping hidden reasoning can narrow that gap inside a speech-native model, while its weaker general-knowledge results show the cost of specializing too aggressively.
Paris-based Kyutai released two speech-native models that take spoken math problems and answer aloud without routing the inference through a transcription system or a separate text language model.
https://x.com/kyutai_labs/status/2102080831242060123
Kyutai announced Voice of Reason in a five-post thread on September 21st. The accompanying research paper, accepted at COLM 2026, is led by Kyutai PhD student Timothee Weisselberger with Edouard Grave and head of research Alexandre Defossez.
The release extends Z.ai's GLM-4-Voice, a model that alternates text and audio tokens as it generates a response. Kyutai changed the post-training rather than the underlying architecture, producing one checkpoint that answers without hidden reasoning tokens and another that silently reasons in chunks while previously generated speech is playing.
The distinction matters for voice systems. Conventional assistants can transcribe a question, send the text to a reasoning model and synthesize its answer. That pipeline gives developers access to stronger text models, but each stage adds latency and strips away some information carried by speech. A native speech model has to begin talking quickly while still doing enough computation to answer correctly.
Two ways to make a voice model think
Kyutai's reasoning-free checkpoint generates its answer directly. The second, STITCH-style checkpoint, inserts unspoken reasoning chunks between segments of the audible response. Because audio takes longer to play than tokens take to generate, the model can work on its next chunk while the user hears the previous sentence.
That design avoids a separate pause for a complete chain of thought. It also means the model's hidden reasoning can continue after the spoken answer has started, an approach intended to preserve the responsiveness expected from a voice interface.
Kyutai trained the models on 150,616 examples derived from Orca-Math. The researchers used Qwen 3 235B to rewrite questions and answers into speech-friendly language, then synthesized the audio with a text-to-speech model and multiple voices.
The direct inference path therefore avoids transcription and a text LLM, but the training pipeline relied heavily on text models and synthetic speech. Reinforcement learning also used a Qwen model as a binary judge of whether an answer was correct. The judge received the question and generated response without the reference answer, making this an AI-feedback system rather than a deterministic arithmetic verifier.
The 77.1% result needs a denominator
Kyutai says the STITCH-style release scored 77.1% on GSM8K, a 1,310-question grade-school math benchmark. The reasoning-free release scored 70.3% under the same decoding setup described in the paper. The original GLM-4-Voice baseline scored 27.3%.
Those headline figures come from released checkpoints evaluated without the paper's original top-k decoding constraint. The main paper reports results averaged across three independent training runs: 65.5% with a standard deviation of 1.1 points for the reasoning-free approach and 74.8% with the same standard deviation for the STITCH-style version.
Kyutai's published materials differ slightly on the first released checkpoint. The announcement thread and paper report 70.3%, while its Hugging Face model card reports 70.6%. The 77.1% figure for the STITCH-style release is consistent across the announcement, paper footnote and model card.
The comparison also stops short of showing speech-native systems at parity with the strongest alternatives. In Kyutai's evaluation, Qwen2.5-Omni reached 84.7%, Qwen3-Omni reached 94.6%, and a cascaded system built around a 31B-parameter Gemma model reached 95.7%. Those systems differ in size and architecture, and two were evaluated through text output, but they establish the remaining accuracy gap.
Kyutai checked whether gains in decoded text survived conversion into actual speech by transcribing the generated audio with Qwen3-ASR. The STITCH-style model reached 72.0% speech-output accuracy after reinforcement learning, up from 66.2% after supervised fine-tuning alone.
Specialization carried a cost. The base GLM-4-Voice scored 40.6% on a spoken subset of TriviaQA. Full-data fine-tuning reduced that result to 33.4% for the reasoning-free model and 20.2% for the STITCH-style version. Reinforcement learning recovered only part of the loss, producing scores of 34.0% and 21.4%, respectively.
Voice of Reason is therefore a math specialist, not a general replacement for a text-backed voice assistant. Kyutai's contribution is narrower and technically useful: reinforcement learning can improve a model that consumes and produces speech directly, while streaming hidden reasoning gives it extra computation without imposing a separate thinking delay. Kyutai has released both checkpoints with instructions for running them through Transformers, though each inherits GLM-4-Voice's license and requires its separate audio front end.