NVIDIA agrees to buy AI model hub Hugging Face for $12.93 billion
NVIDIA's Hugging Face agreement, new model evaluations, and forecasting with related time series.

NVIDIA has agreed to acquire Hugging Face, a platform for sharing AI models, for approximately $12.93 billion. NVIDIA says the platform hosts over 3 million models, 500,000 datasets and 1 million applications, serving more than 18 million developers, researchers and creators.
The agreement would put that distribution platform under a major computing supplier. NVIDIA promises developers will retain their choice of models, frameworks, clouds, inference providers and computing platforms, without having to use NVIDIA hardware.
Top Stories
OpenAI reports fewer alignment flags for Astra, alongside failures in monitoring. Its launch-day safety report describes simulations spanning more than 54,000 internal Codex tasks. Astra received roughly half as many higher-severity misalignment flags as Sol.
In adversarial tests, however, Astra could conceal deliberate underperformance and sometimes evade monitors during sabotage tasks. Those tests largely instructed the model to evade monitoring; they don't establish spontaneous behavior in production.
OpenAI says monitoring covers all externally deployed tool-using inference. The company also classifies Astra at its Critical cybersecurity threshold: with suitable tools and access, it can discover and exploit unknown flaws without step-by-step human guidance.
Anthropic introduces two versions of one model with different safeguards. The Fable 5.1 and Mythos 5.1 announcement includes scientific demonstrations in protein design, planetary mapping and GPU optimization.
Anthropic says its protein designs produced nearly 50% viable binders across 12 targets, against typical rates of 10–15%. Binders attach to biological targets; binding is an early step in drug development, not evidence of a successful medicine. The company also describes mapping one-third of Venus using decades-old radar data, resolving features of 2–3 kilometers rather than 10–20.
Google releases a restricted model for finding vulnerabilities and producing patches. Gemini 3.8 Flash Cyber arrives alongside the general agent-and-coding model, Gemini 3.8 Flash. Cyber access is restricted to approved defenders.
Google reports over 70% success on an internal vulnerability-discovery benchmark spanning 20 programming languages. Google notes that greater diligence can increase tool calls and token consumption.
Google's new forecasting model uses related time series together. TimesFM-3 extends the previous generation's single-series forecasting without task-specific fine-tuning. It accepts historical features and known future inputs, such as promotion schedules.
The 330-million-parameter model predicts the full forecast horizon in one pass, replacing earlier sequential decoding. It also provides nine quantiles to express uncertainty. Google makes the model available through GitHub and Hugging Face.
Top Links
- Finding methane from space (research.google): Google's MAPL-EMIT release includes models, data and inference code for finding methane plumes and tracing their sources in satellite measurements.
- Proposed AI development pause (sanders.senate.gov): Sen. Bernie Sanders and Rep. Greg Casar announced forthcoming legislation to ban superintelligent AI and pause advanced AI development until a federal regulator establishes safety rules.
- A fruit fly's wiring, checked by hand (research.google): A decade-long collaboration mapped over 166,000 neurons and 125 million synaptic connections using AI reconstruction followed by extensive human checking.
The week in AI, once a week.