Key Takeaways
- Samsone, a family of open small audio language models (SALMs) from Samsung R&D Institute Poland and AGH University of Kraków, was introduced in an arXiv paper on September 18, 2026, with its core variant, Samsone-134M, at 134 million parameters.
- Samsone-134M scores roughly 15% above the previous leading SALM on the MMAU benchmark and 36% higher on MMAU-Pro, at 20% fewer parameters than its nearest competitors, though both figures come from the team’s own testing and have not been independently replicated.
- Samsung and AGH University are releasing training code, model weights, mobile-optimised checkpoints and a working Android app, giving researchers a deployable stack rather than just a paper to cite. Samsone-134M, the core model in a new family from Samsung R&D Institute Poland and AGH University of Kraków, scores 36% higher than the previous best small audio model on MMAU-Pro while running in real time on an Android phone. The arXiv paper, posted September 18, 2026, introduces three open models ranging from 99 million to 356 million parameters, and, unusually for a research release, ships a working Android app alongside the weights.
Why Small Models Keep Losing
Until recently, strong audio language understanding meant billion-parameter models on cloud servers. Large audio language models (LALMs) handle tasks like audio question answering well, but they require sending user audio off-device, adding latency and exposing sensitive data to a remote server. For a voice assistant or real-time transcription tool on a phone, that round-trip is often a dealbreaker.
Small audio language models (SALMs) are the compact, locally-run alternative, no network dependency, no data leaving the device. The persistent problem is that compression has historically cost performance. Samsone’s central claim is that the trade-off is less severe than previously assumed.
Architecture: The Separator Token
All three Samsone variants take audio and text as input and produce text output. The audio encoder derives from OpenAI’s Whisper, a transformer built for speech recognition; it extracts acoustic features that are then mapped into the text embedding space via a non-linear modality projector. Samsone-134M uses the SmolLM2-135M language backbone for text generation.
The more interesting architectural piece is a trainable Separator (SEP) token embedding. Most multimodal models struggle when multiple audio clips appear in a single input sequence, the model loses track of where one clip ends and another begins. The SEP token encodes both the count of audio signals and their relative positions, letting Samsone handle interleaved audio and text without that confusion. For conversational AI, where a query might reference several audio segments in sequence, that capability matters.
The Benchmark Numbers
On the Massive Multitask Audio Understanding (MMAU) benchmark, Samsone-134M scores roughly 15% above the previous leading SALM, according to the paper’s results. On MMAU-Pro, the harder evaluation requiring more complex audio reasoning, the gap widens to 36%. Both figures come from the team’s own testing; independent replication has not yet been reported.
The parameter count makes those numbers worth scrutiny. Samsone-134M achieves them at 20% fewer parameters than the prior top models in its class. The paper also reports that Samsone-134M outperforms substantially larger LALMs on these specific tasks, including GAMA and LTU, though MMAU and MMAU-Pro measure a narrow slice of audio understanding, and performance on other tasks may differ.
What the Open Release Actually Includes
Samsung and AGH University are releasing full training code, model weights for all three variants and mobile-optimised checkpoints tuned for the memory and compute limits of edge hardware. There is also a working Android application for real-time on-device inference. Most research releases stop at weights, leaving developers to sort out the deployment stack themselves. A functional Android app substantially lowers the barrier for anyone who wants to test audio AI on real hardware rather than a GPU cluster.
Privacy Gains, Real Limits
Running audio processing locally means voice data never leaves the device. For applications handling ambient audio or personal voice commands, that is a genuine privacy improvement. Latency also drops: on-device inference removes the server round-trip, which matters in interactive settings where a half-second delay is noticeable.
The researchers are candid that compression involves trade-offs. Quantisation and model compression, the techniques that make a 134-million-parameter model fit on a phone, can reduce reliability on audio inputs outside the training distribution. The paper acknowledges limitations without quantifying them precisely, which means developers should expect some performance degradation in noisy or unusual acoustic environments compared to the benchmark numbers. That is an honest position; it is also an incomplete one.
Three Sizes, One Scaling Study
Offering three model sizes, 99M, 134M and 356M, lets the research double as a study of scaling behaviour within the SALM range. The 99M variant targets the tightest hardware constraints; the 356M trades some efficiency for headroom on more complex tasks. Developers can pick based on their device’s neural processing unit capacity rather than treating “small model” as a single fixed point. As chipmakers continue expanding dedicated NPU capabilities on mobile SoCs, the 356M variant that currently pushes limits will likely become the conservative choice.
The 134M results are the paper’s strongest contribution. Whether the 15% and 36% MMAU gains hold across a wider range of real-world audio tasks is what follow-on work will need to establish.
Originally published at https://autonainews.com/samsones-134-million-parameter-model-beats-rivals-in-its-size-class/
Top comments (0)