Editorial illustration for NVIDIA, UMD release AF-Next audio model, beats Phi-4-mm by 12 points on Arabic
NVIDIA's AF-Next Audio Model Beats Phi-4-mm Benchmark
NVIDIA and the University of Maryland released a new audio model that significantly outperforms Microsoft’s on a key benchmark. Their Audio Flamingo Next model scored 21.9 on the CoVoST2 English-to-Arabic speech translation test, beating Microsoft’s Phi-4-mm by 12 points. The model was trained on roughly 108 million audio samples, or one million hours of internet-scale data.
It is also the first to use a “temporal audio chain-of-thought,” linking its reasoning steps to precise timestamps within the audio. The research was fully released.
On LongAudioBench, AF-Next-Instruct achieves 73.9, outperforming both Audio Flamingo 3 (68.6) and the closed-source Gemini 2.5 Pro (60.4). On the speech-inclusive variant (+Speech), AF-Next reaches 81.2 vs. Gemini 2.5 Pro’s 66.2.
The timestamp-based reasoning method addresses a core problem in audio AI: how to parse meaning not just from words, but from their position and context within a longer recording. By making both the model and its massive, open training dataset available, NVIDIA and UMD provide a foundation other researchers can use. The 12-point performance gap on Arabic translation suggests this approach has practical impact.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic NVIDIA NeMo powers telco reasoning model for autonomous network workflows
Industry Applications
Related Topic China offers cheaper electricity to AI firms abandoning NVIDIA chips
Business & Startups
From Archives Developers Claim Measured Drop in Claude's Performance, Sparking Nerf Debate
From Archives Seven AI agents in finance lift cash flow >3% monthly, boost productivity 50%
From Archives NVIDIA launches AITune v0.2.0 with KV‑cache support for LLM inference
Open Source
From Archives Nvidia unveils Agentforce AI platform with Adobe, Salesforce, SAP at GTC 2026
Business & Startups
Common Questions Answered
How does AF-Next improve upon previous audio-language models in Arabic speech translation?
AF-Next demonstrates a significant 12-point improvement over Phi-4-mm, achieving a score of 21.9 in CoVoST2 speech translation for Arabic EN→X translation. This breakthrough represents a major advancement in handling complex audio-to-text translation tasks, particularly for Arabic language processing.
What makes AF-Next unique in the field of large-scale audio-language modeling?
AF-Next is the first fully open audio-language model designed to operate at internet scale, processing approximately 108 million audio samples and 1 million hours of audio. The model introduces a novel Temporal Audio Chain-of-Thought approach that enhances reasoning capabilities for long-form audio content.
What collaborative effort led to the development of the AF-Next audio model?
NVIDIA and researchers from the University of Maryland collaborated to create the Audio Flamingo Next (AF-Next) model, aiming to develop a powerful audio-language system capable of processing raw sound and generating text across multiple languages. This partnership represents a significant step forward in open-source audio-language modeling.
