This week: ServiceNow released Apriel Voice 27B Preview, an open model that listens to speech, reasons about it, and calls tools from one shared brain. There is a free demo you can talk to, and a long list of honest caveats.

Elsewhere: How it works, what the scores do and don't say, how to try it tonight, and three more things that shipped this week.

The New Prime Signal

Most voice assistants are a relay race. One system turns your words into text, a second one figures out what you meant, and a third reads the answer back to you. Every handoff adds delay and a fresh chance to drop the baton. ServiceNow's research team wants to shorten the race. In a blog post dated October 29, it introduced Apriel Voice 27B Preview, an open speech-language model that hears your request, thinks about it, and decides which tool to call, all in the same model. The weights are on Hugging Face under an MIT license, with a technical report on arXiv and a public demo called Apriel Live.

What happened: ServiceNow built Apriel Voice on top of the post-trained Gemma 4 26B-A4B checkpoint. A Whisper large-v3-turbo speech encoder feeds audio into that reasoning backbone through a learned "projector," so speech, text, images, and the results of tool calls all land in the same conversation. According to the model card, it handles transcription, speech translation, spoken question answering, spoken instructions, multi-step requests, and structured function calls. It does not talk back on its own. The model writes text and tool calls, and a separate speech synthesizer turns the reply into audio. ServiceNow says that is deliberate: teams can swap voices or providers without retraining the part that does the thinking. The company calls it a preview with clear limitations and asks demo users to stick to research and evaluation and avoid sharing sensitive information.

Why it matters: A good voice assistant has to do more than spell your words correctly. ServiceNow's own example is someone reporting a locked account: they speak a code, add a screenshot, maybe change their mind halfway, and the assistant still has to take the right next step. Keeping listening and reasoning in one model is meant to make those conversations feel connected instead of passed along. The bigger deal for builders is the license. Help-desk bots, booking assistants, and accessibility tools are exactly the kind of thing companies want to run on their own hardware, and an MIT-licensed model that can hear a request and fire off a tool call is a useful new building block.

The fine print: Every number here is ServiceNow's own, and nobody has independently reproduced them yet. On its English tests, the company reports 65.51% task success on the τ-Voice benchmark with a mean response time of 1.91 seconds, 51.27% on the EVA-A accuracy measure, and 74.13% on EVA-X, which scores conversational experience. Word error rate on LibriSpeech is 2.23% on the clean test set and 5.15% on the harder one. Its own comparison table also shows where it trails: GPT Realtime 2.1 scores higher on EVA-A and τ-Voice, and Gemini-3.8 Live Extended Thinking scores higher on τ-Voice and responds faster. Among the open two-stage systems ServiceNow tested, Apriel Voice comes out on top. The model card warns that tool calls can be wrong, that "faithfulness and reliable completion of complex tool workflows remain limitations," that multilingual performance is uneven, and that noisy or distant-microphone speech is still hard. It also wants 16 kHz mono audio, and ServiceNow's standard setup runs on four GPUs, so this is not a phone app.

This is a refreshingly candid release. ServiceNow put the scores where it loses right next to the ones where it wins, and the limitations section reads like it was written by someone who has actually watched a voice bot misunderstand a customer. It won't replace the polished assistants from the biggest labs tomorrow, and it isn't trying to. What it does is give everyone else an open, inspectable starting point for voice agents that can act, not just chat. And separating the "brain" from the voice is a sensible move. Nobody wants to retrain a 27-billion-parameter model just because marketing picked a new voice.

Think of it as one good listener with a phone book of tools, plus a separate voice actor who reads its replies out loud.

How it works: Training happened in two stages, according to the technical report. Stage one connects the speech interface to the backbone. Stage two adapts the projector and the backbone together using supervised fine-tuning, replayed text data, and teacher-model supervision, with automatically built spoken dialogues as conversation practice. For long recordings, the model card says audio is split into 30-second chunks, and the team saw good performance on continuous recordings up to around 30 minutes, with quality that can slip on longer ones. ServiceNow also built an optimized serving setup that it says roughly doubles output speed in its own measurements. That is the version behind its evaluations and the live demo, where spoken replies come from Cartesia Sonic 3.6.

Try it: The easiest way in is the browser. Open Apriel Live, talk through a small task, and try interrupting it or asking a follow-up. ServiceNow specifically wants to hear where it works and where it doesn't. Just don't read it your bank details. If you have serious GPUs, the model card includes ready-made serving commands and code examples for audio input and tool calling. For the training recipe and the full scoreboard, the technical report has the details.