Apriel Voice 27B Preview
ServiceNow released Apriel Voice 27B Preview, an open speech-language model built on Gemma 4 26B-A4B that handles speech, reasoning and tool calls in one model.
The model takes speech directly and produces both replies and tool calls, so no separate dialogue model is needed, and it still accepts text and image input. It was trained in two stages: first a speech interface, then joint adaptation of the projector and backbone using supervised fine-tuning, text replay, and cross-entropy plus forward KL distillation. The authors say they run it on a custom serving platform for interactive voice use.
- Date
- Thursday 29 October 2026
- Lab
- ServiceNow (Apriel)
- Kind
- model
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Big Bench Audio (thinking mode) | 98.90 | company |
| Speech-BFCL (thinking mode) | 81.13 | company |
| τ-Voice task success rate | 65.51% mean latency 1.91 seconds | company |
| English EVA accuracy | 51.27% | company |
| EVA experience pass@1 | 74.13% | company |
Only the arXiv abstract was read, so the license and weight download location are unconfirmed.
Sources
This record was partly confirmed: some claims could not be checked on 2 November 2026. How we check