Apriel Voice 27B Preview unifies speech, reasoning, tools

Researchers introduce Apriel Voice 27B Preview, an open speech-language model that unifies speech recognition, reasoning, conversation, and tool use within a shared backbone. Built on the Gemma 4 26B-A4B checkpoint, the model directly processes speech and generates responses and tool calls through a two-stage training procedure and efficient online distillation. In thinking mode, it achieves strong performance on benchmarks including Big Bench Audio and Speech-BFCL. Open article →