Model Overview

nvidia/audio-flamingo-next-captioner-hf is the long-form captioning checkpoint in the Audio Flamingo Next family. It is designed for rich, descriptive outputs over long and complex audio, including speech-heavy recordings, environmental sound scenes, and music.

Description

Audio Flamingo Next (AF-Next) is the next-generation open audio-language model in the Audio Flamingo series, built for speech, environmental sound, and music understanding with audio inputs up to 30 minutes.

This checkpoint corresponds to AF-Next-Captioner, the model obtained at the end of AF-Next mid-training after the training mixture is expanded with newly collected long-audio captioning and QA datasets. It is the best AF-Next variant when you want:

• long-form descriptive captions
• verbose summaries of long audio
• timestamp-aware scene breakdowns
• exhaustive descriptions of events, speakers, and background audio

Best For

• detailed audio captions that integrate speech, sound effects, ambience, and music in one response
• prompts that ask for a broad caption while also transcribing spoken content by all speakers precisely
• long-form summaries, timestamp-aware event descriptions, and dense scene breakdowns
• showcase-style qualitative outputs where you want the model to describe how the audio evolves over time

AF-Next Variants

Because this checkpoint comes before the RL-based assistant alignment stage, it is often more verbose and more caption-like than nvidia/audio-flamingo-next-hf. If you want chat-oriented or safer assistant-style behavior, use the instruct checkpoint instead.

These Hub weights are released as an audio-text-to-text model. The broader AF-Next project also discusses streaming TTS and voice-to-voice interaction, but those components are not part of this checkpoint.

This model is for non-commercial research purposes only.

Usage

Install

AF-Next is supported in Transformers:

pip install --upgrade pip
pip install --upgrade transformers accelerate

Notes

• The processor expects mono 16 kHz audio.
• Audio is internally processed in 30-second windows.
• The released processor is configured for up to 1800 seconds of audio, i.e. 30 minutes.
• This checkpoint is best used with prompts that ask for descriptive, coherent, or timestamped captions.
• Prompting matters: this checkpoint is strongest when you ask explicitly for a dense caption, timestamped scene summary, lyrics, or a speaker-aware breakdown.

Prompt Guide

Long-Form Captioning Example

import torch
from transformers import AutoModel, AutoProcessor

model_id = "nvidia/audio-flamingo-next-captioner-hf"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModel.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
).eval()

conversation = [
    [
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": (
                        "Write a detailed caption of this audio. Cover the speakers, "
                        "background sounds, major events, and how the scene changes over time."
                    ),
                },
                {
                    "type": "audio",
                    "path": "https://huggingface.co/datasets/nvidia/AudioSkills/resolve/main/assets/videoplayback_superman.wav",
                },
            ],
        }
    ]
]

batch = processor.apply_chat_template(
    conversation,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
).to(model.device)

if "input_features" in batch:
    batch["input_features"] = batch["input_features"].to(model.dtype)

generated = model.generate(
    **batch,
    max_new_tokens=2048,
    repetition_penalty=1.2,
)

prompt_len = batch["input_ids"].shape[1]
completion = generated[:, prompt_len:]
text = processor.batch_decode(
    completion,
    skip_special_tokens=True,
    clean_up_tokenization_spaces=False,
)[0]

print(text)

Prompting Tip

AF-Next-Captioner responds best to prompts like:

• "Write a rich caption of the full audio."
• "Describe how the audio scene evolves over time."
• "Give a timestamped summary with speaker and sound-event details."

Training Summary

AF-Next-Captioner is the checkpoint produced after AF-Next mid-training. In that stage, the model:

• retains the earlier AF-Next recognition and skill data mixture
• adds newly collected long-audio captioning and QA data
• increases the maximum audio length to 30 minutes
• increases the total context length to 128K
• emphasizes long-form data by down-sampling the earlier stage mixture and assigning full blend weight to long-audio datasets

The broader AF-Next training recipe also includes:

• AudioSkills-XL, LongAudio-XL, AF-Think, AF-Chat, and MF-Skills
• 45K additional multi-talker speech samples
• 200K+ long-form internet videos
• 2M+ real-world short audio skill samples
• 1M multi-audio instruction examples
• 30K multi-turn chat examples
• 386K safety and instruction-following examples
• multilingual ASR and AST data from Emilia, CoVoST, MUST, Amazon-SIFT, ALI Meeting, aidatatang, AISHELL, and Granary
• training on 128 NVIDIA H100 GPUs

Architecture

The released checkpoint exposes AudioFlamingoNextForConditionalGeneration with AudioFlamingoNextProcessor. At a high level, AF-Next combines:

• an AF-Whisper audio encoder using 128-bin log-mel features
• non-overlapping 30-second audio chunking
• a 2-layer MLP audio adaptor
• a Qwen2.5-family text backbone extended to long context
• RoTE for timestamp-aware temporal grounding

The released config uses:

• audio_config.hidden_size = 1280
• audio_config.num_hidden_layers = 32
• text_config.hidden_size = 3584
• text_config.num_hidden_layers = 28
• text_config.max_position_embeddings = 131072

Selected Results

From the AF-Next paper, the captioner-style variant is especially strong on broad open-ended benchmarks:

• MMAU v05.15.25 average: 75.76 for +Captioner
• MMAR: 63.0 for +Captioner
• MMSU: 63.3 for +Captioner

The paper positions this variant as the most caption-oriented AF-Next checkpoint, particularly for long-form descriptive prompting.

Limitations

The paper highlights several limitations:

• internet-scale audio remains noisy and unevenly distributed across domains and languages
• low-resource languages, rare events, and specialized domains are still underrepresented
• very long-context reasoning remains challenging when evidence is distant or sparse
• evaluation does not yet fully cover all AF-Next capabilities, including diarization, timestamped captioning, and voice-to-voice interaction

If you want better assistant alignment, use nvidia/audio-flamingo-next-hf. If you want explicit reasoning traces, use nvidia/audio-flamingo-next-think-hf.

License / Terms of Use

The model is released under the NVIDIA OneWay Noncommercial License. Portions of the dataset generation are also subject to the Qwen Research License and OpenAI's Terms of Use.

Citation

@misc{ghosh2026audioflamingonext,
  title={Audio Flamingo Next: Next-Generation Open Audio-Language Models for Speech, Sound, and Music},
  author={Sreyan Ghosh and Arushi Goel and Kaousheik Jayakumar and Lasha Koroshinadze and Nishit Anand and Zhifeng Kong and Siddharth Gururani and Sang-gil Lee and Jaehyeon Kim and Aya Aljafari and Chao-Han Huck Yang and Sungwon Kim and Ramani Duraiswami and Dinesh Manocha and Mohammad Shoeybi and Bryan Catanzaro and Ming-Yu Liu and Wei Ping},
  year={2026},
  howpublished={Technical report},
  url={https://afnext-umd-nvidia.github.io/}
}