Tencent open-sources AuK, a 1.5B model that edits speech like text

Share
Tencent open-sources AuK, a 1.5B model that edits speech like text

A quiet morning in the open-weights scene, and two releases that both point the same way: the tooling around voice AI is getting strong enough, and cheap enough, that the community no longer needs a closed API to use it.

Tencent open-sourced AuK, a 1.5B-parameter foundation model for speech generation and editing, alongside AuK-Flash, a distilled variant that produces audio in four inference steps. The weights, code and technical report went public under the MIT license, with the paper posted to arXiv on September 8. The model was trained on roughly 3.03 billion instruction-audio pairs and 1.95 million hours of effective supervision, and everything — zero-shot voice cloning from a reference clip, text-to-speech driven by a written voice description, rewriting what was said inside a recording, emotion and paralinguistic edits, noise removal, and separating mixed audio — runs through one natural-language instruction interface rather than five separate models. The Flash variant is where the practical interest sits: distilled down to four generation steps, it runs 4.5 times faster than the full model with no quality-class collapse. The strategic read: Tencent keeps shipping genuinely open weights in voice while most US labs gate their best audio behind paid APIs, and a permissively licensed model that can clone, edit and clean speech in one pass is raw material for dubbing, podcast and agent-voice pipelines that currently rent those features per minute. It follows the real-time audio race we tracked last month — Meta ships Muse Voice Transcribe, its first real-time audio model.


Bartowski — the uploader behind a huge share of the quantized model files people download to run AI locally — rebuilt how those files are made, replacing a fixed recipe with one learned per model. Quantization squeezes a model's weights into fewer bits so a 27B-parameter model fits on a consumer GPU; the catch is that not every layer loses quality equally when squeezed. Bartowski wrote a testing framework with an AI coding assistant, then ran roughly 96 hours of experiments producing and evaluating over 1,000 quantized variants to map which weight groups in a model are actually sensitive. The result, "per-tensor layout maps," bumps the delicate layers to higher precision and crushes the tolerant ones, and the same method validated across other model families. Two details matter beyond the marginal quality gain: a solver decided the layouts rather than a human rule, and every new architecture now faces an automated canary test — one model failed it, and the pipeline correctly fell back to the old method instead of shipping a bad quant. Below about 3 bits per weight, dense models land roughly where they always were, so this is a refinement, not a revolution — but it is what a healthy open-source toolchain looks like: one maintainer's multi-day experiment quietly upgrades everyone's downloads.

What to watch: whether AuK gets independent TTS Arena-style rankings, and how fast the rest of the quantization ecosystem adopts per-tensor layouts.

If a permissively licensed model could clone and edit any voice in four steps, where would you want that capability banned first — and who would enforce it? Write to us via our contact page.

Sources: AuK (GitHub)

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

Minitap says Google's Artemis shipped its code with the credit stripped

Minitap says Google's Artemis shipped its code with the credit stripped

Google's Pixel test team open-sourced an Android automation agent this week. The team it was built on top of says the authorship trail was rewritten before anyone saw it. Minitap says Google's Artemis contains its code, and that the names of the people who wrote it were removed. Minitap published mobile-use in August 2025 under Apache 2.0 — an agent that drives real Android and iOS apps, 2,821 GitHub stars, and the reference implementation behind a clutch of mobile-agent papers.

Hugging Face runs async RL training with no NCCL and no weight ship

Hugging Face runs async RL training with no NCCL and no weight ship

The hardest part of reinforcement learning is not the algorithm. It is getting a trainer and a bunch of inference servers to keep passing weights back and forth without one of them stalling. Hugging Face just published a recipe that does it with three rented jobs, one storage bucket, an HTTP proxy, and no NCCL at all. Hugging Face trained a reasoning model end to end using async GRPO and LoRA, and the whole weight sync is a file upload.

Tencent's EVIE turns a scanned page into something you can actually search

Tencent's EVIE turns a scanned page into something you can actually search

Two releases land in the same place today: Tencent putting open-weight visual document retrieval models on Hugging Face, and vLLM publishing what happens when you run speculative decoding on AMD silicon instead of Nvidia's. One is about what AI can read; the other is about how cheaply you can run it. Tencent has released EVIE-8B and EVIE-4.5B, open-weight models built to retrieve pages of documents by what they look like — and the smaller one nearly matches the big one.

Fastino's 340M decision model beats its own 1B on routing

Fastino's 340M decision model beats its own 1B on routing

Most agent stacks answer a yes/no question by having a large model write a paragraph. Fastino's answer is to not ask a large model at all — and its new release is small enough to run on the CPU you already have. Fastino shipped GLiNER2.5-Decide, a 340M-parameter classifier that takes a bounded question — intent, routing, sentiment, severity, spam — and returns a label in a single forward pass, with no prompt template and no generated tokens.

Huawei open-sources Pangu 2.0's training stack, code and all

Huawei open-sources Pangu 2.0's training stack, code and all

Three things worth your time this hour: the first Chinese lab to give away the code that produces a frontier model, the platforms quietly replacing Hugging Face inside China, and a small OCR model that beat the big generalists on their own benchmark — a month before anyone announced it. Huawei has open-sourced the code that trains Pangu 2.0, not just the weights — publishing the pretraining, supervised fine-tuning and post-training reinforcement-learning pipelines for the model family.