Audio-Flamingo-3-BF16-ComfyUI

Research-only audio understanding, transcription and audio question answering with text output.

This is a converted model package for Jack Spooner’s ComfyUI fork, derived from nvidia/audio-flamingo-3-hf. It is not a new fine-tune.

Required code

The integration modifies ComfyUI core; this is not a drop-in custom_nodes package.

• Model integration commit
• Implementation · Offline converter
• Development branch · Pinned compatible reference snapshot 9d9053c5e
• Installation, node usage and acceptance details

For an existing ComfyUI installation, use the linked core changes and their prerequisites against a compatible base. The integration commit alone is not asserted to be an independently cherry-pickable patch. The pinned reference snapshot includes the required shared loaders, attention and memory changes. Follow the fork’s installation README and dependencies when reproducing that snapshot; update the application only while its queue is idle. Upstream ComfyUI is Comfy-Org/ComfyUI; its README describes the underlying node engine and model management.

Files and installation

Place audio_flamingo_3_bf16.safetensors in ComfyUI/models/text_encoders/. The tokenizer and required inference configuration are embedded; original HF shards are not required at inference.

File	Size
audio_flamingo_3_bf16.safetensors	16.540 GB

Manifest/provenance files record the conversion and source revisions. SHA256SUMS and FILES.json describe this release. Filenames and model folders matter. No original source snapshot needs to be converted again.

Usage

Select audio_flamingo_3_bf16.safetensors in Load CLIP, type audio_flamingo, then connect to Generate Text. Supply one AUDIO clip and an instruction. Start with sampling off, thinking off, and an explicit 512-token output budget. The optional AF-Think adapter and speech-generation weights are not included.

Verification and limits

Short-clip full-weight RTX 5090, transcription/music description, cancellation and unload/reload acceptance is documented. Long-audio accuracy, low-memory operation and other GPU backends remain unverified. The native path consumes audio windows until the actual context limit rather than silently applying the upstream ten-minute crop. Input plus output must fit the model context.

These statements describe the linked retained tests, not a new cross-platform acceptance suite. File size is not a minimum RAM/VRAM requirement. No inference-time downloads or external inference worker are required by these native model paths.

Conversion and provenance

Converts 830 BF16 model tensors to native names, with the original tokenizer embedded. No fine-tuning or precision reduction. This package is the understanding model, without the optional AF-Think adapter or voice generation.

Source model revision: 7d4bae64ee29878af6504ae6f6bb3e40492838ad. See PROVENANCE.json, the conversion manifest and NOTICE. No training data, private inputs or generated user recordings are included.

License and attribution

RESEARCH ONLY. NVIDIA’s OneWay Noncommercial License permits redistribution only with its full license and retained notices; its use limitation defines noncommercial as research purposes only. Original DOCX and a text transcription are included. The upstream card also identifies Qwen Research License and OpenAI terms for portions of dataset generation; those upstream disclosures are preserved and linked.

See LICENSE, NOTICE, and the retained upstream model card/source notices. Model authors and publishers retain credit; this is a third-party ComfyUI repackaging and does not imply endorsement. The weights’ terms and the separately linked ComfyUI code license are distinct.

The underlying Qwen2.5-7B backbone is credited to Qwen/Alibaba Cloud and its Apache-2.0 text is retained in LICENSES/QWEN-BACKBONE-APACHE-2.0.txt. This does not relax the NVIDIA research-only license on this Audio Flamingo derivative.