SleuthVLM: High-Accuracy Video Understanding Without Direct Audio
November 3, 2026
SleuthVLM achieves 77.8% on Video-MME-v2 by using a VLM to process transcribed speech rather than raw audio. This 27B parameter model outperforms Gemini 1.5 Flash and top open-source omni models on long-form video benchmarks.
HOW THIS AFFECTS YOU
●
builderYou can achieve state-of-the-art video understanding by combining VLMs with speech-to-text pipelines.
●
researcherThis challenges the necessity of direct audio-in-video training for omni-modal performance.