Model Overview

Audio Flamingo Next (AF-Next) is the next-generation open large audio-language model in the Audio Flamingo series. nvidia/audio-flamingo-next-hf is the instruction-tuned checkpoint for general audio understanding, question answering, and conversation over speech, environmental sounds, and music.

Description

Compared with Audio Flamingo 3, AF-Next adds:

• a stronger foundational audio-language model for speech, sound, and music
• training data scaled beyond academic benchmarks using public and internet-scale sources
• native long-audio support up to 30 minutes
• stronger multilingual ASR, multi-talker speech understanding, and long-form captioning
• timestamp-aware modeling through Rotary Time Embeddings (RoTE)

This checkpoint corresponds to AF-Next-Instruct, the post-trained assistant variant. It is the best default checkpoint if you want:

• general audio QA
• instruction following
• multi-turn audio chat
• long-form audio understanding
• timestamp-aware prompts

Best For

• standard audio QA and instruction following across speech, sound, and music
• assistant-style responses for long-audio questions, follow-up questions, and multi-turn chat
• speech understanding tasks such as ASR, paralinguistic understanding, and multilingual AST / speech translation
• music captioning and broad audio description when you want a direct answer instead of a dense long-form caption

AF-Next Variants

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.
• Prompting matters: this checkpoint is strongest when the task and output format are explicit. Ask directly for QA, ASR, AST, timestamps, or speaker labels instead of using a generic prompt.

Prompt Guide

Single-Turn Audio + Text

import torch
from transformers import AutoModel, AutoProcessor

model_id = "nvidia/audio-flamingo-next-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": (
                        "Transcribe the speech, identify important background sounds, "
                        "and mention approximate timestamps for key events."
                    ),
                },
                {
                    "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=1024,
    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)

Multi-Turn Follow-Up

conversation = [
    [
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": (
                        "Transcribe the input speech, then keep the conversation open "
                        "for follow-up questions."
                    ),
                },
                {
                    "type": "audio",
                    "path": "https://huggingface.co/datasets/nvidia/AudioSkills/resolve/main/assets/WhDJDIviAOg_120_10.mp3",
                },
            ],
        },
        {
            "role": "assistant",
            "content": [
                {
                    "type": "text",
                    "text": "Summer follows spring, the days grow longer, and the nights are warm.",
                }
            ],
        },
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": "Summarize the content in one sentence.",
                }
            ],
        },
    ]
]

Training Summary

AF-Next is trained with a four-stage curriculum spanning pre-training, mid-training, post-training, and temporally grounded reasoning training. The paper describes:

• AF-Whisper-based audio modeling with broader multilingual and multi-talker speech coverage
• expanded training data from AudioSkills-XL, LongAudio-XL, AF-Think, AF-Chat, and MF-Skills
• 45K additional multi-talker speech samples
• 200K+ long-form internet videos spanning roughly 5 to 30 minutes
• 2M+ real-world short audio skill samples mined from long-form audio
• 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

AF-Next-Instruct is obtained after GRPO-based post-training focused on multi-turn chat, safety, instruction following, and selected AudioSkills-XL skills.

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

Limitations

The paper highlights several limitations:

• internet-scale audio is still noisy and uneven across domains, languages, and acoustic conditions
• low-resource languages, rare sound events, and specialized domains remain underrepresented
• long-context reasoning is still difficult when relevant evidence is sparse or far apart in time
• evaluation does not yet fully cover all supported capabilities, including multi-talker ASR, diarization, timestamped captioning, and voice-to-voice interaction

For most users, this is the best AF-Next checkpoint to start with. If you need explicit long-form reasoning traces, use nvidia/audio-flamingo-next-think-hf. If you want the most verbose descriptive captions, use nvidia/audio-flamingo-next-captioner-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/}
}