mlx-community/AuK-Flash-bf16

tencent/AuK-Flash (Tencent Hunyuan; technical report arXiv:2609.08936) converted to MLX for xocialize/mlx-auk-swift.

AuK is a unified speech generation and editing model. One natural-language instruction covers:

zero-shot and instruct TTS;
content edits (replace / insert / delete words);
acoustic edits (pitch, speed, volume);
paralinguistic edits (emotion, timbre, de-accent, nonverbal, whisper);
enhancement and separation.

Flash: the distilled 4-step recipe, CFG off (4 DiT forwards). RTF ≈ 0.13 bf16 on an M5 Max.

⚠️ Licence: read before use

The DiT and VAE in this repo are MIT (Tencent's LICENSE, included). AuK requires a semantic encoder, the Qwen2.5-Omni-3B thinker, and that encoder is under the Qwen RESEARCH LICENSE AGREEMENT: research and evaluation only, non-commercial. A pipeline that runs is therefore research-only, whatever this repo's own badge says. The encoder is published separately with its licence: mlx-community/Qwen2.5-Omni-3B-Thinker-bf16 (also -8bit and -4bit). Tencent has said it plans a permissive encoder (Tencent-Hunyuan/AuK#13).

Files

file	contents
`dit.safetensors`	Flux2Edit DiT (10 MMDiT + 20 single-stream blocks, dim 1536) plus the 36-layer fusion weights. bf16; the fusion logits, scale and RoPE frequencies are fp32.
`vae.safetensors`	BigVGAN-Flow VAE (24 kHz, hop 480, 64-dim latents). bf16 with fp32 normalisation statistics and anti-alias taps. Weight-norm folded; the training-only coupling flow removed.
`config.json`	Architecture, VAE and sampling configuration.

Conv weights are in MLX (O, K, I) layout. Key names are upstream's verbatim. Converted from tencent/AuK-Flash@575b92f0 with Tools/convert/convert.py. A V1 gate checks it: the fp32 conversion equals upstream's folded weights bit for bit.

Parity

The reference is the upstream PyTorch code (Tencent-Hunyuan/AuK@9481474) run in fp32. Measured on six cookbook tasks with the same noise, on Base and Flash:

the Swift port's generated latent is within relL2 ≤ 6.6e-4 at fp32, and the waveform within ≤ 7.5e-3;
bf16 against fp32 over 36 renders: error rate −1.2 points, CAM++ SIM −0.023, DNSMOS −0.01;
the 8-bit and 4-bit encoders pass the same gate.

Receipts: PORTING-SPEC and MEASUREMENTS.

Use (Swift)

import AuKCore
let pipe = try await AukPipeline.load(modelDir: aukDir, encoderDir: thinkerDir, dtype: .bfloat16)
let wav = try pipe.generate(AukRequest(instruction: AukInstruction.emotion("happy").text(),
                                       audio: samples, sampleRate: 24000, seed: 1))