ServiceNow published Apriel Voice 27B Preview on Hugging Face in late October 2026. It is an open-weight model that listens to speech and replies with text and tool calls, under the MIT license. It is built for voice agents: the support line that books, cancels or looks things up for you. Unlike full voice-to-voice models, it doesn't speak. You add a text-to-speech system for the spoken reply.
What Shipped
| Item | Detail |
|---|---|
| Architecture | Whisper large-v3-turbo audio encoder, a learned projector, and a Gemma 4 26B-A4B language backbone |
| Size | 27B total parameters, BF16 |
| Input | Audio (16 kHz mono), text and images |
| Output | Text and tool calls. No native speech output and no native full-duplex |
| License | MIT |
| Languages | Tagged for 49. The model card says quality is uneven outside English |
| Long audio | Processed in 30-second chunks. Tested to about 30 minutes, up to an hour without guarantees |
| Serving | vLLM, tensor parallel across 4 GPUs. The optimized setup was tested on 4 H100 80 GB |
Reported Benchmarks
All of these are ServiceNow's own figures from the model card.
| Benchmark | Score |
|---|---|
| LibriSpeech word error rate, clean / other | 2.23% / 5.15% |
| τ-Voice pass@1 | 65.51, with 1.91 s mean latency including speech synthesis |
| Big Bench Audio | 98.90 |
| MMLU-Pro | 84.42 |
| GPQA Diamond | 80.40 |
Coverage notes it trails OpenAI's GPT Realtime 2.1 on some voice scores.
Why It Matters
Open voice-agent models lower the cost of putting an AI on a phone line or in an app. Under MIT, any company can build one on Apriel without paying per call to a closed provider. Expect more voice agents in customer service, booking and ordering, run by businesses you deal with every day.
What It Means for Brand Visibility
A voice agent answers in one spoken reply, with no list of links and no second option on screen. When a caller asks "which plan should I get" or "can you recommend a plumber", the agent names one or two things, often by calling a tool. That makes two things matter: whether your brand is reachable as a tool the agent can call, and whether the model behind it knows your brand well enough to say it. Speech recognition matters too. Brand names that are hard to say or easily misheard are more likely to be mangled on the way in. See voice AI brand visibility and the October 2026 release roundup.