tencent

/

AuK

Like 343

Follow

Tencent 12.3k

Thanks for releasing AuK under MIT — the editing results are the strongest in the open, and the unified instruction interface is exactly what downstream products need. One licensing gap is worth stating on the card, because it makes the release non-commercial in practice.

The gap. AuK's own weights and code are MIT (LICENSE), but inference cannot run without Qwen/Qwen2.5-Omni-3B, which the README has users download separately — and that model is under the Qwen RESEARCH LICENSE AGREEMENT: §1.i defines "Non-Commercial" as research or evaluation purposes only, §2.a grants rights FOR NON-COMMERCIAL PURPOSES ONLY, and §2.b requires a separate license from Alibaba Cloud for commercial use.

Why adopters cannot route around it. Every task's condition passes through that encoder: CFMEdit.encode_text fuses all 36 thinker layers with the learned layer_weights (shape [36] in auk_base.safetensors) and projects them through txt_proj (2048 → 1536), so the checkpoint is bound to the 3B's 36 × 2048 geometry. The Apache-2.0 siblings do not fit — Qwen2.5-Omni-7B is 28 × 3584 and Qwen3-Omni is 48 × 2048 — and re-conditioning the backbone is the pre-training run (50k + 600k updates on 256 GPUs per the report), not the released fine-tune. Net effect: for commercial products the whole pipeline is non-commercial despite the MIT badge, and license reviews that check every weight file in the inference graph (ours does) reject it outright. We evaluated AuK this week for an on-device Apple Silicon (MLX / Swift) port and had to pass on this point alone; the architecture itself would be a clean port.

Requests — full detail, code permalinks and acceptance criteria are in Tencent-Hunyuan/AuK#13:

1. A variant trained on a permissively licensed encoder (Apache-2.0 / MIT). A smaller encoder would be ideal for on-device use, and even a somewhat lower-quality variant would unblock commercial adoption entirely.
2. An explicit note of the encoder dependency and its license in the License section of this card and of tencent/AuK-Flash, so adopters are not surprised after integrating.
3. If a Tencent–Alibaba Cloud arrangement extends commercial rights for Qwen2.5-Omni-3B to AuK users, documenting it here would resolve this without any retraining.

Happy to test and evaluate a permissive-encoder release, including on Apple Silicon (MLX), and to report parity and benchmark numbers back.