Samsone: Open Audio-Language Models for Smartphones

Samsung released Samsone — a family of three open small audio-language models (99M, 134M, and 356M parameters) for on-device deployment, with the 134M version setting a class record on the MMAU benchmark at 61.33%.

Compiled by Sergey KostenchukPublished 2026-09-22Updated 2026-09-22

2026-09-22 Research HuggingFace

🤖 Samsone: Open Audio Models for Smartphones

Samsung released Samsone — a family of three open audio models (SALM) for on-device deployment: 99M, 134M, and 356M parameters. The 134M model outperforms systems 20–60 times larger on the MMAU benchmark: 61.33% versus Mellow's 53.34%, and higher than Audio Flamingo 2 (3B) and Qwen2-Audio (8.4B).

🌍 Speech and sound models have become so compact that speech, music, and sound understanding can be moved from the cloud to the smartphone — for privacy and low latency.

👤 Try it today: Python API, CLI for .wav files, and an Android app (API 31+). Limitations: responses only in lowercase English — a targeted sound analysis tool, not a universal assistant.

Source 1: https://arxiv.org/abs/2609.21666 Source 2: https://github.com/SamsungLabs/samsone

Sources