AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing
Paper • 2609.08936 • Published • 206
AuK — MAESTRO mirror
Verbatim mirror of tencent/AuK (MIT), re-laid out into
per-variant lanes for MAESTRO's downloader. No weights are modified — files are byte-identical
to upstream.
| Lane | Contents | Size |
|---|---|---|
base/ |
auk_base.safetensors (1.53B DiT, fp32) + vae.safetensors (BigVGAN-flow, 24 kHz) + config.yaml |
6.76 GB |
⚠️ This model does not run on its own
The AuK checkpoint contains no text encoder — zero text_encoder.* tensors. It requires a
Qwen/Qwen2.5-Omni-3B snapshot at runtime, which is
not mirrored here and is not MIT:
Qwen RESEARCH LICENSE — grants use "FOR NON-COMMERCIAL PURPOSES ONLY", where non-commercial means "for research or evaluation purposes only".
MAESTRO downloads that encoder on demand, directly from Qwen's own repository, so the end user is the licensee. In practice this makes AuK a non-commercial model despite AuK's own MIT terms.
There is no permissive substitute: config.yaml pins text_hidden_dim: 2048 to the 3B thinker's
hidden size, and the AuK layer-fusion head is a [36] vector — exactly one weight per
text_config.num_hidden_layers. The Apache-2.0 Qwen2.5-Omni-7B (3584-dim) is not a drop-in,
and upstream states Qwen3-Omni is unsupported.
Licenses & attribution
- AuK code and weights — MIT, © 2026 Tencent. See
base/LICENSE. - Qwen2.5-Omni-3B encoder — Qwen RESEARCH LICENSE (non-commercial), © Alibaba Cloud. Built with Qwen.
Notes measured on this checkpoint
- Output is 24 kHz mono; latent dim 64, downsample 480 (50 latent fps).
- The DiT is 1.5305B params, shipped fp32. bf16 measures 1.59 dB log-mel from fp32 — 0.21× the sampler's own seed-to-seed variance — while running 2–3× faster, so bf16 is the default in MAESTRO.
- Peak VRAM ~8.4–9.0 GiB on a 12 GB card, achieved by never keeping the 7.52 GiB bf16 encoder and the DiT resident at the same time.