Samsone-134M: Samsung’s open audio model for on-device specs

September 26, 2026

arXiv.org Poland

Overview

Researchers from Samsung R&D Institute Poland and AGH University of Kraków have introduced “Samsone,” a family of open Small Audio Language Models (SALMs) designed for on-device inference. Its core model, Samsone-134M, established new state-of-the-art (SOTA) performance across multiple benchmarks within its size class. This model employs a multimodal LM architecture that accepts both audio and language inputs to generate text outputs, significantly expanding the feasibility of advanced AI applications in resource-constrained environments.

In Depth

Key Findings

A research team from Samsung R&D Institute Poland and AGH University of Kraków has announced “Samsone,” a family of open Small Audio Language Models (SALMs) specifically optimized for on-device inference. The core model of this family, Samsone-134M, has achieved new state-of-the-art (SOTA) performance across multiple audio-language benchmarks, despite its compact size, significantly elevating the capabilities of small models.

Technical / Clinical Details

The Samsone model exhibits the following key technical features:

Multimodal LM Architecture: Samsone adopts a multimodal Large Language Model (LLM) architecture capable of integrally processing both audio and text inputs. This enables users to interact with AI by combining voice commands, environmental sounds, or text prompts.
Optimized for On-Device Inference: With a compact model size of 134 million parameters, direct inference is possible on resource-constrained edge devices such as smartphones, wearables, and IoT devices. This allows for real-time AI processing without transmitting data to the cloud, enhancing privacy protection and reducing latency.
Achievement of SOTA Performance: Samsone-134M has demonstrated results surpassing existing best models within its size class across audio understanding, text generation, and cross-modal tasks. This proves that high-quality AI functionalities can be delivered even with small models.
Open Source: The model’s open-source release will allow a broad range of researchers and developers to rapidly build new applications and improvements upon Samsone as a foundation.

These characteristics position Samsone to potentially revolutionize various edge AI applications, including voice assistants, smart home devices, and in-car infotainment systems.

Background & Context

While LLMs have achieved high performance in recent years, their massive size and computational costs have hindered deployment on edge devices and in privacy-sensitive environments. Small Language Models (SLMs) have emerged as a promising approach to address this challenge, but balancing performance and efficiency has been difficult. Samsone’s introduction overcomes this trade-off, paving the way for deploying high-quality AI functionalities on edge devices. The focus of research institutions from major corporations like Samsung in this area strongly suggests significant growth in the future edge AI market.

Strategic Significance & Outlook

On-device SALMs like Samsone are critically important from the perspectives of privacy, security, and latency. The ability to process complex audio and text directly on a device, without reliance on cloud-based AI services, will dramatically enhance the user experience. The future outlook includes accelerated development of new multimodal applications built upon Samsone, leading to the proliferation of more personalized and responsive smart devices. Furthermore, as other research institutions and companies also focus on developing similar small models, innovation in edge AI is expected to accelerate even further.

Source: https://arxiv.org/html/2609.21666v1

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.