NVIDIA + UMD released AF-Next: open audio-language model that outperforms Gemini-2.5-Pro on MMAU-Pro (75.01% vs 57.4%). Temporal Audio Chain-of-Thought anchors reasoning to timestamps.