Benchgen

OmniBench — Results

RankModelScore
1lfm3-omni-2b46.32
O

OmniBench

1 phaseActive

A tri-modal benchmark testing simultaneous visual, acoustic, and textual reasoning in universal omni-language models.

Overview

OmniBench

Category Metric Saturation Created

Paper

Quick answer: OmniBench ("Towards The Future of Universal Omni-Language Models") is a benchmark introduced by Li, Ma, Zhang et al. (submitted Sep 2024, last revised v6 Dec 2025) to test whether multimodal large language models can concurrently process and reason about visual, acoustic, and textual inputs together, rather than handling modalities one at a time. It targets genuinely omni-modal models rather than vision-language-only or audio-only systems.

What OmniBench Measures

Most multimodal benchmarks test one extra modality layered onto text (image+text, or audio+text). OmniBench instead requires simultaneous tri-modal reasoning — a model must process visual, acoustic, and textual information together to answer correctly, which the authors argue better reflects the demands on emerging "universal omni-language models" that take any combination of text, image, video, and audio as input.

Public Benchmark Scores

ModelOverallSpeechSourceDate
LFM3-Omni-2B46.32%48.25%Liquid AI model card2026-10-27
  • LFM3-Omni-2B — currently the only Benchgen-tracked model with a public OmniBench score.

Sources

  • Li, Yizhi, Ma, Yinghao, Zhang, Ge, et al. "OmniBench: Towards The Future of Universal Omni-Language Models." arXiv:2409.15272, submitted 2024-09-23, last revised 2025-12-29 (v6).