ServiceNow's open-weight speech model, uploaded to HuggingFace on October 13, 2026 as a preview. It adds a speech interface to Google's Gemma 4 26B-A4B: a Whisper large-v3-turbo encoder feeds the Gemma 4 backbone through a learned projector. Training first aligns the speech interface, then adapts it together with the backbone on speech tasks with text replay, using cross-entropy or teacher-distillation supervision depending on the data source. The model takes audio, text and images and emits text and tool calls. It does not speak: voice applications pair it with an external synthesizer, and the card's evaluations use Cartesia Sonic 3.5. MIT license, about 50 languages tagged on the card.

ServiceNow's reported results, all self-run with thinking mode on: EVA-A 51.27, EVA-X 74.13 and τ-Voice pass@1 65.51 at 1.91 s mean latency, against GPT Realtime 2.1 (59.03 / 61.6 / 69.82) and a Parakeet + Gemma 4-26B cascade (39.97 / 3.93 / 67.78). So it beats the open cascades on the two EVA scores but not on τ-Voice task success, and trails the proprietary speech-to-speech systems on accuracy. LibriSpeech WER is 2.23 (test-clean) and 5.15 (test-other). Text reasoning stays close to the base model: MMLU-Pro 84.42 against Gemma 4's 85.27, GPQA Diamond 80.40 against 77.80, and identical AIME 2024/2025 scores. The card reports good results on recordings up to about 30 minutes.

Model Details

Architecture MOE
License MIT
Base model gemma-4

Benchmark Scores

Benchmark Score Mode
EVA-A 51.27 —
EVA-X 74.13 —
τ-Voice pass@1 65.51 —
LibriSpeech test-clean WER 2.23 —
GPQA Diamond 80.40 text input, thinking
open-weightspeechmultimodalagentic

Related