Introduction

AMBIT is now a runnable open checkpoint, not just a paper. The team (HKUST / Noiz AI / University of Washington) published WJM-George/AMBIT on Hugging Face on 2026-11-04 (created ~16:56 UTC), with matching inference code on GitHub and headphone demos on the project page.

One MIT model handles text-guided First-Order Ambisonics (FOA) generation and instruction-guided FOA editing. Editing also takes a reference FOA clip. That is a rare day-0 surface for makers who care about where a bark or a car pass sits in a 3D soundfield — not only what it sounds like in mono.

What Shipped

  • Vendor / product: AMBIT — Kam Man Wu, Yu Zhang, Pengjun Fang, Mingfei Chen, Weijia Chen, Zeyue Tian, Qifeng Chen (HKUST / Noiz AI / UW)
  • Modality: generative + editable spatial audio (native FOA)
  • Open vs closed: open weights and code under MIT (bundled Qwen text / planner components keep their own license files under LICENSES/)
  • Runnable today: hf download WJM-George/AMBIT + python -m ambit.infer from github.com/WJM-George/AMBIT on Linux + CUDA with bfloat16; listen first at wjm-george.github.io/AMBIT
  • Concrete fact: shared Transfusion backbone (AR ScenePlan + rectified-flow render); edit/reference FOA is 44.1 kHz, four-channel ACN/SN3D in W, Y, Z, X order; root model.safetensors alone is about 1.52 GB, with a separate FOA VAE (595 MB) and text encoder (1.40 GB) in the same repo

Run It

git clone https://github.com/WJM-George/AMBIT.git
cd AMBIT
python3 -m venv .venv && source .venv/bin/activate
python -m pip install -e .
hf download WJM-George/AMBIT --local-dir models/AMBIT

python -m ambit.infer generate --bundle models/AMBIT \
  --prompt "A dog barks to the left while a car passes from front to right." \
  --output outputs/generated.wav --seed 42

python -m ambit.infer edit --bundle models/AMBIT \
  --source input_foa.wav \
  --prompt "Move the dog to the right while keeping the other sounds unchanged." \
  --output outputs/edited.wav --seed 42

Both paths write FOA audio plus a generated ScenePlan (source content, activity, 3D trajectories). The paper framing is arXiv:2610.34415 — Spatial Audio Generation and Editing with Executable Plans and Self-Distilling Transfusion — but the November 4 HF bundle is the zero-day maker artifact.

Limits Worth Knowing

  • Card requires Python 3.10+, Linux, and a CUDA GPU with bfloat16 — not a laptop-CPU toy.
  • Editing expects proper FOA references in the documented channel order; stereo WAVs are the wrong input shape.
  • Spatial demos are meant for headphones; mono laptop speakers will flatten the point of the release.
  • Third-party text/planner stacks inside the download keep their own licenses — read LICENSES/ before commercial redistribution.
  • Distinct from Stable Audio-style mono/stereo text-to-audio already covered elsewhere: AMBIT’s product claim is native FOA gen and edit in one backbone.

Original Source

Conclusion

If you need open, instructionable First-Order Ambisonics — generate a moving scene from a prompt, then nudge one source without flattening the rest — the November 4 AMBIT checkpoint is the card to pull.

—Anabel ♡