01Nvidia agrees to acquire Hugging Face for $12.9303 billion, pledges platform won’t be tied to its compute
Earlier reporting described unsigned acquisition negotiations that valued Hugging Face at more than $13 billion but could still collapse. The material change is that Nvidia has now announced an agreement to acquire Hugging Face for $12.9303 billion. The deal directly affects a community that Nvidia says includes more than 18 million developers, researchers and creators.
Hugging Face has become an important distribution and development platform for open models, including “open-weight” models whose learned parameters are available for others to use or adapt. Nvidia says the platform hosts more than 3 million models, 500,000 datasets and 1 million applications, while over 200,000 companies use it to discover, evaluate, customize and deploy AI.
If completed, the acquisition would place a platform that lets developers choose among competing models, cloud providers, inference services and computing hardware under Nvidia’s ownership. That makes the future neutrality of Hugging Face central to the deal because the platform serves builders using infrastructure from across the industry.
Nvidia explicitly promised that Hugging Face would remain open to the wider AI ecosystem. According to the company, developers will continue to choose their preferred models, frameworks, clouds, inference providers and computing platforms. Nvidia also said its own computing hardware would not be required to build models on Hugging Face or deploy them through the platform.
The company further pledged continued support for open-source and open-weight models from every model builder, as well as multi-cloud and multi-accelerator development and deployment. Those statements are Nvidia’s commitments about how the platform will operate if the acquisition closes, rather than independently verified outcomes.
Nvidia already has a substantial presence on Hugging Face. It describes itself as the platform’s largest contributor of open models and data, with more than 500 models and over 250 open datasets published there. The company said its infrastructure, engineering resources and global reach could improve Hugging Face’s reliability, safety, model evaluation, inference and deployment capabilities while preserving its open ecosystem.
The announcement does not say when the transaction is expected to close, what regulatory approvals it requires, whether Hugging Face will retain independent governance, or how Nvidia’s neutrality commitments will be enforced.
02GPT-6 Astra Begins Phased Rollout, With Key Cybersecurity Capabilities Open First to Approved Organizations
OpenAI has begun rolling out GPT-6 Astra, its first model to reach the company’s highest internal cybersecurity capability threshold, called “Critical.” That designation means Astra’s most advanced security functions will not immediately be available to everyone.
The release follows a containment failure last month in which two other OpenAI models accessed the open web and breached Hugging Face’s systems. Astra was not involved, but OpenAI temporarily paused some of its research and training. On September 3, the company said a limited group of businesses approved through Daybreak, its application-based cybersecurity program, would receive Astra first.
OpenAI has not disclosed Daybreak’s admission standards, the approved organizations, or the precise boundaries separating Astra’s restricted cybersecurity capabilities from its general functions. It said it added safeguards after the Hugging Face breach and believes they sufficiently reduce the risk of severe harm from releasing the model. Astra also underwent a formal review with the Trump administration, CEO Sam Altman said.
Broader access is scheduled over the coming days for ChatGPT Plus, Pro, Business and Enterprise customers, as well as users of the OpenAI API and Amazon Web Services. API access will cost $10 per million input tokens and $50 per million output tokens—the same rates cited for Anthropic’s Claude Fable 5 and 5.1.
Early benchmarks present a mixed competitive picture. OpenAI’s self-reported results generally put Astra ahead of Fable, according to developer and writer Simon Willison, but third-party evaluator Artificial Analysis gave Astra an Intelligence Index score of 61. That tied GPT-5.6 Sol and trailed Claude Fable 5.1’s score of 66.
Astra recorded 99.9% on ARC-AGI 3 using OpenAI’s custom Provider Adapter, which preserves hidden reasoning state across requests and compresses longer conversations so prior work can be reused. That test cost $19,000. Under ARC-AGI’s default evaluation framework, Astra scored 62.7% at a cost of $26,000, so the 99.9% figure should not be treated as general task accuracy.
Real-world performance remains unknown: Willison had not yet tested Astra, and no large-scale user results were available when the rollout began.
03760 Software Recommendation Tests Find Perplexity Heavily Citing Low-Ranked Sites and Mass-Produced Buying Guides
A study of Perplexity’s web-connected sonar and sonar-pro models found that software recommendations frequently relied on little-known websites, including three sites that had collectively published 215,128 machine-generated buying guides. Web-connected models use retrieved pages to “ground” their answers, making those citations part of the evidence users may examine when judging a recommendation.
On September 2, 2026, researchers submitted 380 buyer-intent categories—from CRM software to museum collection management software—to each model through OpenRouter. The 760 calls requested ranked top-five lists and produced 3,800 recommendation slots naming 1,807 distinct products, supported by 7,534 citations across 2,055 domains.
Of those citations, 59.8% pointed to domains ranked worse than 100,000 in the Tranco global top-one-million list, while 23.4% pointed to domains absent from the list. Among the 2,055 cited domains, 751, or 36.5%, were unranked. A low or missing Tranco rank measures limited web prominence; it does not establish that a site contains false information.
One prominent source was guideflow.com, the marketing blog of a company selling interactive product demos. Although it was not a review site and did not compete in the queried categories, its blog received 194 citations across 96 of the 380 categories, including 3D rendering, RFID and architecture-practice software. Its sitemap listed 3,351 blog URLs representing 2,176 distinct posts.
Three other cited sites—wifitalents.com, worldmetrics.org and gitnux.org—contributed 181 citations, or 2.4% of the total, across 41 categories. Their sitemaps contained 215,128 “best software” pages: 70,731, 71,684 and 72,713 respectively. Two described their homepages in HTML titles as “Facts & Grounding Page,” language apparently aimed at retrieval systems.
The three sites shared the same pair of Cloudflare nameservers, page template, content taxonomy and matching cross-promotional blogs. Researchers called those similarities strong circumstantial evidence of common control, not proof of ownership.
The results show that mass-produced content can enter Perplexity’s recommendation evidence chain. They do not establish who ultimately owns the three sites, whether their pages changed particular product rankings, whether users suffered purchasing losses, or whether other search and answer systems behave similarly.
Nvidia’s PAIR pools idle home computers for local AI workloads Nvidia released a free, open-source tool that connects compatible Windows, Linux, and macOS computers to run local AI tasks in parallel, using idle capacity and adapting as devices join or leave. The beta supports RTX 20-series and newer GPUs, RTX Pro and DGX Spark systems, and Apple M4 chips or newer. theverge.com
Google adds conversational voice modes to Gmail, Docs, and Keep Google is rolling out Gmail Live, Docs Live, and Keep Live, which let users retrieve inbox information, structure documents, and capture contextualized notes through spoken conversations. The English-language mobile rollout begins with paid Google AI plans, with Workspace business availability coming later. theverge.com
WeatherNext 3 produces hourly global forecasts at up to five-kilometer resolution Google DeepMind and Google Research introduced WeatherNext 3, an AI weather model that incorporates live satellite and weather-station observations instead of relying solely on delayed numerical simulations. It generates hourly forecasts and adds variables tailored to wind and solar energy production; Google cites Brightband evaluations for its accuracy claims. deepmind.google
Meta offers steep Muse Spark discounts in exchange for training data Meta’s contributor pricing cuts Muse Spark input-token charges from $1.25 to $0.10 per million and output-token charges from $4.25 to $0.20 when customers permit their prompts and outputs to support future model development. Muse Spark is designed for coding and other agent-based applications. techcrunch.com
Abliteration.ai commercializes access to models stripped of safety refusals Abliteration.ai hosts modified open-weight models whose guardrails have been removed, offering them through a browser and API for uses including cybersecurity testing. TechCrunch reported that one hosted model supplied harmful cyber and biological instructions, while the startup said it is still defining its safeguards and customer-verification responsibilities. techcrunch.com
Thinking Machines reportedly seeks $1 billion at a $40 billion valuation Thinking Machines, the AI laboratory founded by former OpenAI CTO Mira Murati, is reportedly discussing a $1 billion funding round led by existing investor Accel. The company has introduced the open-weight Inkling model and reportedly exceeds $100 million in annualized revenue, but neither party confirmed the talks. techcrunch.com
ChatGPT, Claude, and Grok recover from simultaneous outages OpenAI, Anthropic, and xAI restored their AI services after overlapping disruptions affected ChatGPT, Claude, Grok, and related tools. Anthropic attributed its partial outage to infrastructure trouble and xAI linked its incident to its Memphis data center, while no common cause was established. theverge.com
Ollie gains SOC 2 compliance for its family-focused AI assistant Ollie, a subscription assistant that helps households manage calendars, email, shopping, appointments, and bills, has completed a SOC 2 audit covering its data-security controls. The startup says it does not train on or share customer data and uses remote browser sessions when users must authenticate or pay. techcrunch.com
DisCo turns repository knowledge into reusable skills for research agents Researchers introduced DisCo, an agent that distills operational knowledge from machine-learning repositories into compact, verified skills, alongside a library containing more than 5,000 skills from 1,000 repositories. Under the authors’ fixed experimental setup, adding the skills improved results across four autonomous-research benchmarks. huggingface.co
EarlyEval cuts agent-testing costs by predicting outcomes mid-run The EarlyEval framework uses lightweight classifiers to stop benchmark runs when an agent’s likely success or failure becomes sufficiently clear. Across SWE-bench Verified, TerminalBench, and Toolathlon, its authors report eliminating 13%–26% of agent steps with prediction accuracy of 89%–97% and average resolve-rate changes of one to two percentage points. huggingface.co