Releases·Model Release
Mizar: A 159M-Parameter Audio-Language Model for Audio Understanding
Researchers introduce Mizar, a 159.3M-parameter audio-language model built from a CED-Small audio encoder connected to SmolLM2-135M via a frequency-merging mapper. A three-stage training recipe on ReasonAQA, AudioMCQ, and AVQA data pushes it past all prior ALMs under 200M parameters on MMAU, MMAR, and ADQA-clean. Code and checkpoints are public, and the model runs on-device with roughly one-second CPU latency.
0 engagement1 sourceWed, Sep 23, 2026, 04:21 PM
AI Overview
- Mizar pairs a CED-Small audio encoder with SmolLM2-135M via a frequency-merging mapper, trained in three stages on ReasonAQA, AudioMCQ, and AVQA
- Across five seeds, Mizar reaches 52.92% on MMAU, 42.42% on MMAR, and 36.02% on ADQA-clean, beating all ALMs under 200M parameters
- Mizar runs local CPU inference at 1.09-second mean latency on MMAU questions; code and checkpoints are on GitHub
Related
Releases·Model Release
Samsone: A Family of Open Small Audio Language Models for On-Device Inference
A new arXiv paper introduces Samsone, a family of open Small Audio Language Models (SALMs) for edge computing, with sizes of 134M, 99M, and 356M parameters. Samsone-134M sets a state-of-the-art for its size class while remaining competitive with far larger models. The team releases training code, weights, mobile-optimized checkpoints, and an open-source Android app for real-time on-device inference.
A new arXiv paper introduces Samsone, a family of open Small Audio Language Models (SALMs) for edge computing, with sizes of 134M, 99M, and 356M parameters. Samsone-134M sets a state-of-the-art for its size class while remaining competitive with far larger models. The team releases training code, weights, mobile-optimized checkpoints, and an open-source Android app for real-time on-device inference.
0 engagement1 sourcearxivFri, Sep 18, 2026, 12:03 PM
Releases·Model Release
SNU researcher releases Sori-1B, a 1B audio-language model trained from scratch without text-only pretraining
Sori-1B is a 1B-parameter audio-grounded language model released on Hugging Face by a single SNU researcher. Its decoder is trained entirely from scratch on audio-paired text, with no text-only pretraining or pretrained-LM initialization, so answers ground in audio rather than text-only priors. The release includes evidence that typical audio models lean heavily on those priors.
Sori-1B is a 1B-parameter audio-grounded language model released on Hugging Face by a single SNU researcher. Its decoder is trained entirely from scratch on audio-paired text, with no text-only pretraining or pretrained-LM initialization, so answers ground in audio rather than text-only priors. The release includes evidence that typical audio models lean heavily on those priors.
6 engagement1 sourceredditMon, Aug 31, 2026, 02:48 AM
Developments·Paper
EvoAudio: arxiv paper presents recursive self-improvement loop for audio understanding, lifting scores up to 6.3 points
An arXiv paper introduces EvoAudio, a recursive self-improvement system for audio language models. It uses model performance to generate progressively harder, verifiable training data via audio tools, updated with reinforcement learning across 13 evolution rounds. The approach raises performance up to 6.3 points on audio reasoning benchmarks without new human annotation.
An arXiv paper introduces EvoAudio, a recursive self-improvement system for audio language models. It uses model performance to generate progressively harder, verifiable training data via audio tools, updated with reinforcement learning across 13 evolution rounds. The approach raises performance up to 6.3 points on audio reasoning benchmarks without new human annotation.
0 engagement1 sourcearxivWed, Sep 23, 2026, 05:45 AM
Developments·Paper
Tracing Audio Grounding and Answer Selection in Audio LLMs
A new arxiv paper examines what must change inside Audio LLMs for the provided audio, rather than textual cues or linguistic priors, to determine the model's answer. The authors study the common remedy of training on data whose answers cannot be inferred from text alone, which improves performance but leaves the internal mechanism unclear. The work matters to practitioners building audio understanding models who need answers genuinely grounded in audio.
A new arxiv paper examines what must change inside Audio LLMs for the provided audio, rather than textual cues or linguistic priors, to determine the model's answer. The authors study the common remedy of training on data whose answers cannot be inferred from text alone, which improves performance but leaves the internal mechanism unclear. The work matters to practitioners building audio understanding models who need answers genuinely grounded in audio.
0 engagement1 sourcearxivFri, Sep 4, 2026, 02:16 AM
Developments·Paper
FireRedAudio introduces unified audio language model with decoupled representations
FireRedAudio is a general-purpose audio language model that uses decoupled continuous representations to unify understanding and generation tasks. It is built on a shared 9B-parameter LLM and claims to be the first publicly disclosed model of its kind. This matters for practitioners as it addresses key representation challenges in audio AI.
FireRedAudio is a general-purpose audio language model that uses decoupled continuous representations to unify understanding and generation tasks. It is built on a shared 9B-parameter LLM and claims to be the first publicly disclosed model of its kind. This matters for practitioners as it addresses key representation challenges in audio AI.
0 engagement1 sourcearxivTue, Aug 25, 2026, 07:37 AM