ServiceNow's voice agent model: It connects the Whisper large-v3-turbo speech encoder to Gemma 4 26B-A4B via a projection layer, directly listening to spoken requests to perform transcription, speech translation, and speech question-answering. It can also reason and issue tool calls in OpenAI format while retaining text and image inputs. It only outputs text and tool calls, does not produce audio itself, and requires external speech synthesis to function as a voice assistant. Weights were uploaded on 10 29, and it was not included in previous issues.
Upvotes 65 · Likes 65 · Downloads 338
Reasons
- Officially claimed performance on voice customer service benchmarks: EVA-A (service accuracy) 51.27, EVA-X (dialogue experience) 74.13, and τ-Voice task success rate 65.51. The entire system has an average response time of 1.91 seconds (with thinking enabled and Cartesia Sonic 3.5 synthesis). In the same table, the open-source three-stage solution of Parakeet + Gemma 4-26B + TTS scores 39.97, 3.93, and 67.78; GPT Realtime 2.1 scores 59.03, 61.6, and 69.82.
- Text capabilities remain largely unaffected after adding voice: Officially claimed scores for text input are MMLU-Pro 84.42 (base model 85.27), GPQA Diamond 80.40 (base model 77.80), and τ² 70.95 (base model 68.93). English speech recognition word error rates on LibriSpeech test-clean / test-other are 2.23% / 5.15%.
- Long audio is automatically segmented into 30-second chunks. Officially claimed stable performance for continuous recordings of around 30 minutes; service examples are configured for 30 minutes, and parameters for 1 hours are also provided. The model card lists 49 language tags.
Usage
- vLLM: Use the Docker image provided in the model card, amant555/vllm_apriel_voice:plain-26b-v0.32.0 (or PR #64683 from upstream vLLM). Run
vllm serve --model ServiceNow-AI/Apriel-Voice-27B-preview --trust-remote-code --tensor-parallel-size 4 --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4to start an OpenAI-compatible interface; the example also uses google/gemma-4-26B-A4B-it-assistant for speculative decoding. There is also an optimized image, which officially claims approximately double the output throughput on 4 H100 80GB GPUs. - When calling, include an
audio_urland a text instruction in the messages. For transcription and translation, use temperature 0.4 and disable thinking (enable_thinking=falseinchat_template_kwargs). For general instructions and tool tasks, use temperature 1.0, top_p 0.95, top_k 64, and enable thinking. Audio must be 16 kHz mono. - If you do not want to deploy, you can try it on the official Apriel Live page (servicenow-ai-apriel.hf.space); for offline inference with Transformers, set
trust_remote_code=True.
Risks
- All results are self-reported by the official team: EVA / τ-Voice benchmarks use the official optimized image and Cartesia synthesis, and latency measures the entire system, which will vary with different deployments. The official team notes that multilingual performance is uneven; English scores do not represent other languages, and no separate scores are provided for Chinese.
- BF16 weights are approximately 54.5 GB, and the official example uses a 4-card tensor parallel configuration; the Docker image is published under the personal account amant555, and native vLLM support is still in the PR stage.
- Preview version: The official team warns that tool calls may fail, complex multi-step workflows may not be reliably completed, noisy or far-field recordings are difficult to recognize, and long recordings may result in repetition or topic drift. Human oversight is required for operations with real-world consequences.