Quiet Day, Loud Moves: The Stack v3, FLUX 3 Goes Physical, and a $10.3B Bet on Inference

The biggest open code library ever

Hugging Face released The Stack v3, now the largest openly published code dataset in the world: 114 TB of raw code drawn from 224 million repositories across 770 programming languages, filtered down to roughly 5 trillion training tokens. That is nearly a tenfold jump from the previous version's ~550 billion tokens, with huge gains in C++, TypeScript, Rust and Python. Because it ships freely, it lets any lab, university or startup train competitive coding models without depending on a closed ecosystem, and it lands right in the middle of a heated policy debate about whether open models and open data are a strategic risk.

FLUX 3 crosses from pictures into the physical world

Black Forest Labs launched FLUX 3, a single unified architecture spanning image, video, audio and action prediction, with early access to FLUX 3 Video and an explicit path toward robotics. Robotics firm mimic immediately built FLUX-mimic, a video-action model trained on robot and wearable data that runs on a single on-prem GPU and is already being tested with Audi. The bet: better video world-modelling transfers directly into better robot control.

A new voice for Qwen

Alibaba's Qwen team shipped Qwen-Audio-3.0-TTS in Flash and Plus variants, covering 16 languages with inline control tags like [whisper] and [angry], natural-language style steering, robustness to noisy references, and up to three minutes of speech in a single pass. It claims the top spot on a major independent text-to-speech leaderboard. A smaller experimental system also appeared offering per-word control over duration, loudness, pitch and tone.

$300M to run the world's inference

Chip startup Etched raised a $300M Series C at a $10.3B valuation and opened an 80,000 square foot, 10 megawatt facility. Its pitch is deliberately narrow: not training frontier models, but running inference, the everyday work of answering your questions, faster and cheaper than general-purpose hardware.

Health arrives in ChatGPT

OpenAI rolled out Health in ChatGPT in the United States, letting users connect Apple Health and supported medical records. The company says that data gets additional encryption, is not used to train foundation models or target ads, and was reviewed heavily by physicians. It is less a new model than a higher-trust application layer, alongside a quieter desktop voice-control rollout.

Trillion-scale training, tuned like a pit crew

The vLLM project detailed prime-rl 0.6.0, a serving and reinforcement-learning stack using FP8, expert parallelism, prefill/decode disaggregation and KV offload to train a trillion-scale model on coding tasks with sub-five-minute steps across 28 H200 nodes, and it is open.

Editing a model's mind by hand

A researcher claims to have written around 500 Wikipedia facts directly into the weights of an 8-billion-parameter Llama model with no fine-tuning, LoRA or RAG, and built a visual map of where each fact physically lives. It raises real questions about side effects, but hints at a new form of persistent memory.

Austria goes sovereign and open

Austria rolled out a government AI platform for around 180,000 federal employees, built on open Mistral models and an open-source interface, running on its own data centres, a concrete public-sector proof that open weights enable private, swappable, in-house AI.