AuK-Flash and AuK: Tencent Quietly Released an Open Speech Model That Generates, Edits, and Separates Audio

Tencent released two of the most consequential open speech models of 2026 this week and told almost nobody. Around September 7, the Tencent organization on Hugging Face made public AuK — "a 1.5B foundation model for speech generation and editing," in its own model card's words — and AuK-Flash, the distilled four-step variant of the same model, both under an MIT license. There is no Tencent blog post, no social announcement and no press release attached to either repository; between them the two have been downloaded a few dozen times, the code repository the cards link to is not reachable, and no independent benchmark exists anywhere on the indexed web. This piece lays out what is actually knowable from the two repositories, the configuration files and Tencent's own project page — and what is still unverified.

The claim is worth taking seriously even without independent numbers, because the scope is unusual. AuK is not another voice-cloning model. Through a single natural-language instruction interface it promises zero-shot and instruction-based text-to-speech, editing that reaches into what is said, how it is said and the non-verbal sounds around it, plus speech enhancement, speech separation and music separation. The field has historically built a different specialist model for each of those jobs. Tencent's bet, in one generative backbone, is that a single model can do all of them — and that AuK-Flash can do it fast enough for real-time-ish use.

Two checkpoints, one release

The Hugging Face API timestamps tell a story of a repository that sat private before it shipped. The base tencent/AuK repository was created on August 18, 2026 and tencent/AuK-Flash on August 21; both were last modified on September 8, and the model card carries a news entry dated September 7 — "[2026/09/07] AuK is now open-source. Code and model weights are publicly available." The natural reading is that Tencent built the repositories in mid-to-late August, flipped them public around September 7, and touched the card again on September 8. Nothing about the model appears anywhere on the indexed web, and download counters sit in the tens — so "quiet release" is not a rhetorical choice, it is the literal situation.

The two checkpoints are the same architecture at different points in its life. AuK is the teacher: a base model aimed at high-quality generation. AuK-Flash is the student: distilled for fast four-step inference, which is what makes it the more interesting deployable target and the reason it leads the title here. On Tencent's project page the model family is presented as roughly 1.5 billion parameters, trained on what the card describes only as "millions of hours of diverse audio data" and the project page specifies as about 1.95 million hours of effective supervision across roughly 3.03 billion instruction–audio instances. The project page also carries Shanghai Jiao Tong University and Tencent Hunyuan marks, placing AuK inside Tencent Hunyuan's open-model research line rather than in a commercial product family. Training is not the only thing that happened after pre-training: Tencent describes a post-training stage that mixes human-feedback preference optimization for open-ended editing with reward-based reinforcement learning for speech generation.

The variant table in the card is unambiguous about which is which: AuK is "the base model for high-quality generation," AuK-Flash is "the distilled model for fast 4-step inference." Both are downloadable from Hugging Face and from ModelScope under the Tencent-Hunyuan namespace, and both carry the MIT license — among the most permissive licenses an open-weights release can have, which matters for anyone wanting to build a product on top.

What one instruction-driven model actually does

AuK's interface is a single prompt box: you hand the model a natural-language instruction and, where a task needs one, an audio reference, and it returns audio. Grouped by the card's own taxonomy, the five task families are:

• Speech generation — zero-shot TTS that clones a voice from a reference clip, and "instruct TTS" that synthesizes speech from a voice description alone ("a warm female voice, slightly slower"), with no reference audio at all.

• Content editing — rewriting what is said: replacing, inserting or removing words in an existing recording, and lyric editing that rewrites the words of a singing clip while keeping the melody and the singer's voice.

• Acoustic editing — pitch in semitones, speaking rate, and volume in decibels, each adjustable on an existing clip while the identity and content are preserved.

• Paralinguistic editing — emotion, timbre (to a description), de-accenting, adding or removing non-verbal sounds like breaths, laughs and coughs, and converting between normal speech and whisper.

• Enhancement and separation — speech enhancement (denoising and dereverberation), speech separation by talking order, music separation that pulls the singing voice out of a mix, and target-speaker extraction identified by what the speaker says.

That list is the actual differentiator. Most capable open speech systems in 2026 are strong at one or two of these — voice cloning and TTS are the crowded corner — and weak or absent on the rest. AuK's wager is breadth: the same weights, the same instruction format, across generation, editing and separation, which is the kind of scope that normally only ships as a collection of fine-tuned specialist models.

How it is built

Reading the repositories and the project page together gives a consistent architecture picture. AuK is not an autoregressive audio language model of the kind that reads text and emits tokens. It is a flow-matching diffusion system in the image-model mould, with three parts:

• A multimodal large language model — Qwen/Qwen2.5-Omni-3B, downloaded separately — reads the instruction and any reference audio and produces the semantic conditioning.

• An audio VAE jointly trained on speech, general audio and music encodes and decodes at 24 kHz through 50 Hz acoustic latents; the repository's config identifies it as a BigVGANFlowVAE with a latent dimension of 64 and a 480-fold downsample, and ships it as a separate vae.safetensors file.

• The generative core is a hybrid rectified-flow Transformer: dual-stream MMDiT blocks that keep text and audio conditioning separate, followed by unified single-stream DiT blocks that mix them, trained under a flow-matching objective. The released config shows a 1,536-wide model with 24 heads, 10 dual-stream layers and 20 single-stream layers.

AuK-Flash is that same transformer after distillation. Tencent describes a two-stage process — trajectory-level consistency initialization, then a "task-routing decoupled" DMD objective — finished with clean-prediction regression supervision. The result, per the project page, is near-teacher generation quality in four sampling steps with no classifier-free guidance at inference, at roughly a 4.5× wall-clock speedup over the full model under matched conditions. Those are vendor claims, unreproduced, but they are the concrete reason AuK-Flash is the variant a builder would reach for.

What Tencent claims, and what a reader can check

The project page makes strong qualitative claims — AuK is "leading" on generation and editing and "competitive" on signal-level restoration tasks like enhancement and separation — and lists the benchmark suites behind them: Seed-TTS-Eval and InstructTTSEval for generation, MMAE-Speech, SpeechEditBench and Ming-Freeform-Audio-Edit for editing, DNSMOS and UTMOS for perceptual quality, and DNS Challenge, CHiME-4 and Libri2Mix for enhancement and separation. No numeric scores are published on the page, there is no paper behind it, and no third party has reproduced any of it. Every one of those claims is vendor-reported and should be read that way until an independent run exists.

Two other claims need flagging because they did not survive a direct check on September 9, 2026. First, the card says code is publicly available and links github.com/Tencent-Hunyuan/AuK for installation, Gradio, ComfyUI and fine-tuning instructions — but that repository returns HTTP 404 as of this writing, and it is not among the Tencent-Hunyuan organization's public repositories. The code may simply still be rolling out; either way, "code is available" is currently the card's claim, not something a reader can act on. Second, the card points to demo Spaces on Hugging Face and ModelScope: the ModelScope studio responds, but the Hugging Face Space returned an authorization wall when checked, so the demos are at best partially public.

What's here, what's missing

Taking stock on September 9, 2026 separates a real release from a complete one.

• Present — two downloadable weight sets under an MIT license: tencent/AuK for the base model and tencent/AuK-Flash for the distilled four-step variant, each about 6.1 GB for the diffusion transformer plus a bundled VAE, mirrored on Hugging Face and ModelScope.

• Present — an unusually broad task surface under one instruction interface: generation, content and acoustic and paralinguistic editing, enhancement, and separation, all in a single model family.

• Present — a concrete "fast" story: AuK-Flash distills to four steps with no classifier-free guidance, at a vendor-claimed 4.5× wall-clock speedup.

• Missing — an announcement. No Tencent blog post, social post, press release or Hunyuan channel post yet points at either repository.

• Missing — a paper or technical report. The citation block on the project page is a placeholder with no authors and no arXiv identifier, and there are no published numbers behind the benchmark claims.

• Missing — accessible code. The GitHub repository and Cookbook the card links return 404 as of the state check, so today the only runnable artifact is the weights plus whatever the community assembles around them.

• Missing — serving and independent evaluation. No inference provider hosts AuK, no hosted API exists, and with downloads in the dozens, nobody has published a reproduction yet.

Running AuK-Flash today, or waiting

If you want to run AuK-Flash yourself, the download path is explicit even though the code is not: pull tencent/AuK-Flash and tencent/AuK (the base is optional unless you want the teacher), and also pull Qwen/Qwen2.5-Omni-3B, because the checkpoint contains only the diffusion-transformer and layer-fusion weights — the multimodal encoder and the VAE load from separate files at runtime, and the card notes that missing text_encoder keys during loading are expected. Hardware is the real constraint. This is a ~1.5B diffusion transformer whose checkpoint alone is ~6.1 GB, running alongside a ~3B multimodal encoder and a VAE, so a 24 GB GPU is the sensible starting point rather than a laptop model; the config's memory notes, which mention gradient checkpointing bringing training peaks from roughly 91 GB down to 75 GB, are about training and should not be read as inference requirements. On software, the config defaults to the flash-attention backend but the inference CLI can fall back to a plain torch attention backend, which removes the flash-attn build step for a first run.

Whether that is worth doing today depends on what you are building. The honest state of play is that this is a model to watch and to test in a sandbox, not one to wire into a production pipeline yet — with no paper, no accessible code and no independent evaluation, the capability claims rest entirely on Tencent's word. For a team that wants speech generation, editing or separation behind an API today, a model that is actually served is the pragmatic choice, which is precisely the role OrcaRouter plays across the models that do have endpoints: one API for 200+ models with provider list prices passed through at zero markup, so a vendor's price is your price from the first day, and automatic failover so an unproven newcomer can be tried on a slice of real traffic without betting the pipeline on it. None of that applies to AuK or AuK-Flash yet, because nothing serves them. The moment a provider hosts AuK, router economics become the cheap way to evaluate it; until then the choice is self-hosting the weights or waiting.

Frequently asked questions

Is AuK or AuK-Flash available through an API?

No. Both are weights-only releases. No inference provider hosts them and no hosted API exists, so there is nothing to call today — the only way to use them is to download the weights and run them yourself.

What is the difference between AuK and AuK-Flash?

The same model family at two stages. AuK is the base model, aimed at high-quality generation; AuK-Flash is its distilled variant, tuned for fast four-step inference with no classifier-free guidance, which Tencent says runs about 4.5× faster than the base under matched conditions. That speedup figure is vendor-reported and has not been independently measured.

Can AuK-Flash be used commercially?

The weights are released under the MIT license, which permits commercial use with the license notice preserved — the least restrictive common open-weights license. The separate code repository that the card links is not reachable as of September 9, 2026, so the practical answer for code reuse is still open until it appears.

What to watch

• The code repository — github.com/Tencent-Hunyuan/AuK is the single highest-value missing artifact. When it appears, the Cookbook's instruction templates turn the task list above from marketing into something runnable.

• An independent run — the first third-party reproduction will settle whether the "leading" generation and editing claims survive contact with real audio, and whether four-step AuK-Flash really approaches the base model.

• A paper — the project page's citation is a placeholder, which usually means an arXiv submission is coming; the training recipe, the benchmark numbers and the distillation detail all live there.

• A serving listing — the first inference provider to host AuK turns it from a directory of weights into a callable model with a price, which is when it becomes trialable through a router rather than a self-hosting project.

AuK and AuK-Flash are real but unfinished in the way the most interesting open releases usually are: the weights are honest, and the story around them has not been written yet. Bookmark tencent/AuK-Flash, treat every capability claim as vendor-reported until a paper or an independent benchmark lands, and check whether the next few days bring the code that would turn this from a promising artifact into a buildable one.