Samsung has released Samsone — a family of three open small audio language models: Samsone-99M, Samsone-134M, and Samsone-356M, designed to run directly on-device. The mid-sized 134M-parameter model set a record for the compact class on the MMAU benchmark, outperforming audio systems 20–60 times larger. Weights, training code, and an Android app are published openly, so speech, music, and environmental sound understanding works offline, and user recordings never leave the smartphone.
What happened
The Samsung team, consisting of Piotr Masztalski, Michał K. Grzeszczyk, and Olaf Sikorski, released the Samsone family of small audio language models: Samsone-99M, Samsone-134M, and Samsone-356M. The architecture is built from off-the-shelf components: a 9M-parameter Whisper-Tiny audio encoder, a nonlinear projector, and Hugging Face's SmolLM2 language model with 135M or 360M parameters. Each audio clip is compressed to 50 audio tokens with a SEP separator, so sound can be inserted anywhere in a text prompt. According to the authors, Samsone-134M scored 61.33% on the MMAU benchmark, compared to 53.34% for the previous compact-segment leader Mellow, outperforming the 3B-parameter Audio Flamingo 2 (61.06%) and the 8.4B-parameter Qwen2-Audio-Instruct (57.4%). On a Galaxy S25 Ultra, the model generates 87 tokens per second without hardware optimizations. The paper on the models was accepted to the Interspeech 2026 conference, which will be held in Sydney from September 27 to October 1.
Context
The class of small audio language models, abbreviated SALM, emerged as a response to cloud-based audio systems: describing sound, distinguishing music, and analyzing environmental noise previously required either sending recordings to a server or models with billions of parameters. The benchmark for the compact segment was Mellow, while the level of large systems was set by Audio Flamingo 2 and Qwen2-Audio. The MMAU and MMAU-Pro benchmarks measure understanding of speech, music, and environmental sounds, with MMAU-Pro built on more complex tasks involving deep audio reasoning. Most models in this class require a separate audio prefix and consume the context of a small language block, whereas Samsone's compact token representation of audio allows freely mixing sound and text in a single prompt. The family's models were trained only on public datasets ReasonAQA and AudioSkills, and the open training code makes the result reproducible and transferable to other small language models.
Why this matters for the industry
The main signal for the industry is that audio understanding no longer requires billions of parameters and cloud access. Audio features can be embedded in mobile products without cost-per-call, without network latency, and without sending user sound outside, changing the economics of features that previously only made sense as cloud APIs. A fully open stack — PyTorch weights, training code, ExecuTorch export, and a ready-made Android app — turns research into a building block: a product with speech, music, and environmental sound understanding can be built today. Training on public data and an open methodology give the community a rare reference point for scaling-law research on compact SALMs — most industrial work rarely reaches this level of transparency. In the next six months, a wave of forks, quantization, measurements on mid-range devices, and community fine-tuning for other languages and domains are likely, and comparison with Samsone on MMAU and MMAU-Pro may become a standard line in tables of new compact models. On a two-year horizon, if the efficiency trend continues, on-device audio understanding could become a standard layer of smartphones, with the cloud remaining for complex cases.
Why this matters for users
Samsone can be tried immediately, and no server is needed. A Python API of the form from samsone import Samsone134M automatically downloads the checkpoint from a GitHub Release, CLI commands samsone134M and samsone356M accept .wav files and describe or compare sound, and the Android app runs on any device with API 31+. Everything runs locally: recordings do not leave the phone, and responses arrive without network latency. A realistic scenario for a developer or enthusiast is to build a prototype of a narrow audio feature in one to two days, such as offline captions for audio notes or clip comparison, and measure latency on target devices. It is worth remembering that this is a point tool: the model responds only in English in lowercase and is designed for describing and comparing sound, not for dialogue as a universal assistant.
What is still unknown / limitations
The record is confirmed on exactly one benchmark — MMAU, and it is incorrect to extrapolate it to audio understanding in general. On the more complex MMAU-Pro, Samsone-134M scored 37.57%, compared to 27.5% for Mellow, but significantly lagged behind Qwen2-Audio (45.41%) and GPT-4o Audio (52.5%): deep audio reasoning remains the domain of large models. Generation speed without hardware optimizations was measured by the authors only on the flagship Galaxy S25 Ultra; there are no public measurements for mid-range devices. According to the authors, after fine-tuning on audio, the model loses general language skills. Multilingualism and vocabulary expansion look like a logical next step, but they are not stated in the authors' plans — scenarios with community ports and product integrations remain community expectations, not Samsung announcements.
Sources
- Samsone: A Family of Open Small Audio Language Models for On-Device Inference (arXiv:2609.21666)
- SamsungLabs/samsone — official project repository on GitHub with weights, training code, and Android app
- Interspeech 2026 — conference where the work was accepted (Sydney, September 27 – October 1)
Author
Look at AI, editorial team