Meta: Muse Spark Open is out, the strongest US open-weight model yet — tuned for Macs together with Apple. On Thursday Meta made the weights of Muse Spark Open downloadable, meaning the trained model’s parameters, which anyone can run on their own hardware. According to the company, it is the most capable US open-weight model to date: it has 816 billion parameters, 42 billion of which are active per token (a so-called mixture-of-experts architecture), a one-million-token context window, and it understands text, images, audio and video. It was designed for coding and long-running, multi-step agentic tasks. Working with Apple, Meta optimized it to run on two linked Mac Studios with M5 Ultra chips; according to CNBC, two to four machines can be linked, and a suitable Mac Studio costs at least $10,799. The model is also available from several cloud providers, and the license allows limited commercial use. According to CNBC, the collaboration is also notable because the two companies were rivals for years, and their relationship hit a low point in 2021 after Apple’s privacy changes. Source: Meta, CNBC, Hugging Face.
The US consumer protection authority has subpoenaed OpenAI and Anthropic over incidents caused by agents. According to Axios, on Wednesday the FTC (Federal Trade Commission) sent the two companies binding information demands of more than 50 pages; an agency spokesperson confirmed this to CBS News. The FTC is requesting documents on recent security incidents — in which AI agents broke into websites and government systems during testing — as well as on the companies’ safety measures, privacy rules and marketing materials that may have misled consumers. According to Axios’s source, the agency also wants to question executives and will later send subpoenas to third parties affected by the agents. We wrote earlier about the FTC opening an investigation in September; this is the first concrete step. According to a WIRED analysis on Thursday, the costs of dealing with damage caused by agents have so far typically been borne by the victims themselves. Source: Axios, Engadget, WIRED.
Anthropic: METR’s independent evaluators have been given “staff-level” access. Anthropic announced on Thursday that six staff members of METR, the nonprofit that evaluates models, have been given company devices, office space and roughly the same internal access as the company’s own risk assessment teams; the exceptions are customer data, information covered by attorney-client privilege and model architecture. METR may publish its analyses of the company’s risks, incidents and practices: Anthropic cannot veto them and can only request redactions in narrowly defined categories, and METR can publicly assess whether a redaction was justified. The company is building a three-tier system: SecureBio carries out narrower reviews and needs approval to publish, Accenture is also an evaluation partner, and talks are under way with Hugging Face. Anthropic compares the arrangement to the supervision of big banks, where the regulator’s experts work inside the bank. We wrote earlier that the labs wanted a say in what outside evaluators could publish. Source: Anthropic, TechCrunch.
Anthropic: investors are talking about a lower IPO valuation, and the slide in AI stocks deepened. According to four Reuters sources, some of the investors who attended Wednesday’s investor day are now talking about a valuation of roughly $1.5 trillion for the IPO expected in November, instead of the roughly $2 trillion expected earlier. Anthropic grew its third-quarter revenue by 45% to $16.8 billion and was profitable on an adjusted basis, but some investors were disappointed by slowing growth: between the first and second quarters, revenue had still more than doubled. According to Reuters, a cooler reception could affect the whole AI market. Meanwhile, the slide in AI-linked stocks deepened: according to The Guardian, South Korea’s Kospi 200 index, dominated by memory chipmakers, plunged more than 8% (SK Hynix nearly 11%, Samsung 9.5%), Japan’s Nikkei fell 4.9% and SoftBank, which also has a stake in OpenAI, fell 10%, helped along by higher-than-expected US inflation of 3.7% in September. Source: Reuters (Investing.com), The Guardian.
Microsoft: MAI-Base-2, the company’s in-house 2.8-trillion-parameter base model, is complete. Microsoft AI said on Thursday that it had finished pre-training its new base model: the mixture-of-experts model has 2.8 trillion parameters, 80 billion of them active. According to the company, on its own non-public tests it beats every open-weight model examined, and used compute 4–12 times more efficiently than the strongest open base model, DeepSeek-V4-Pro-Base. The model was trained on 16,000 Nvidia GB200 chips over 38 days in August and September, without distillation according to Microsoft, meaning it did not learn from other models’ outputs. The model is now in the post-training (fine-tuning) phase; the company would license the weights to trusted partners through the Foundry platform. According to Mustafa Suleyman, head of Microsoft AI, this gives them a frontier in-house base model in the 3-trillion-parameter class too, and post-training results are coming soon. Source: Microsoft AI, Mustafa Suleyman (X).
Open models: Nvidia’s new Nemotron family and Chinese StepFun’s 600-billion-parameter model are available to download. Nvidia released Nemotron 3.5 Super VL, a 120-billion-parameter (12 billion active) open model that understands text, images, documents and video and, according to the company, is built for long-running enterprise agents; according to independent measurements by Artificial Analysis, it scored twice as high as its predecessor on the intelligence index. Alongside it came Nemotron 3.5 Bolt 74B-A9B, which fits into 24 GB of memory and runs agentic tasks on local machines, such as Nvidia RTX/DGX Spark computers; an even smaller 4 GB version arrives on October 23. On Thursday, as promised in September, China’s StepFun published the weights of its Step 5 Preview model on Hugging Face and ModelScope (we wrote earlier about its API): the 600-billion-parameter model, which uses 27 billion active parameters per token, has a one-million-token context window and image understanding and is built for long agentic tasks. Source: Nvidia, Artificial Analysis (X), StepFun (X), Hugging Face.
Financial Stability Board: AI agents need more human oversight. On Thursday the Financial Stability Board (FSB), which brings together the financial authorities of the G20 countries, finalized its recommendations for banks and other financial firms on the responsible adoption of AI. According to Reuters, the final text is stricter on AI agents than the June draft: higher-risk uses require more intensive human oversight, and firms must set clear, risk-based thresholds for when an agent has to ask for human approval. The recommendation to treat AI as a “synthetic employee” was dropped; instead, the text states that agents do not think like humans and may interpret the goals set for them in unexpected ways. Source: FSB, Reuters (Investing.com).
Meta: from October 30, business account users agree to let Meta collect their websites’ content and use it for AI training too. According to Nieman Lab, under Meta’s new commercial terms, businesses using Facebook, Instagram, Messenger or WhatsApp business accounts — practically every news publisher among them — give Meta permission to collect, with automated programs, the content of websites “associated” with their accounts and to use it to develop and train its products; Meta’s products include its AI services. The terms take effect on October 30, and there is no opt-out: according to Meta, anyone who does not accept them must stop using its products for business purposes. Lawyers who spoke to the publication said this effectively amounts to authorization for AI training for publishers; one said a court might not consider it valid, but publishers should assume it would. Source: Nieman Lab.
Perplexity: it has open-sourced SPACE, its system for isolating agents. Perplexity has made the “sandbox” platform that runs all Perplexity Computer tasks available to anyone under the Apache 2.0 license: each agentic task gets its own isolated virtual machine, and credentials and outbound traffic are controlled from outside the virtual machine. The company told Semafor that it instructed nine frontier models to break out, and none of 118 attempts succeeded. Cloud providers CoreWeave, Akamai and Crusoe have committed to offering the system to their customers. The move follows a series of incidents involving agents escaping their test environments. Source: Semafor, Perplexity (X).
OpenAI: new voices and better pronunciation in ChatGPT’s voice mode — but the model got worse on some safety measures. On Thursday OpenAI began rolling out GPT-Live-1.1, the model behind the voice mode in ChatGPT, Codex and dots, and its smaller version, GPT-Live-1.1 Quip: 13 new voices have arrived, including in French, German, Hindi, Mexican Spanish, Korean and Japanese; according to the company, non-English pronunciation has improved a lot, and the model less often says “let me check…”. The update is reaching all users and will come to the API later. According to the model’s safety card, on some synthetic tests GPT-Live-1.1 performed measurably worse than its predecessor in the sexual, mental health, emotional reliance and gore categories; OpenAI says it is still investigating why and will fix this in future versions, though these measurements do not account for system-level protections such as crisis lines and parental controls. Source: OpenAI (X), OpenAI (release notes), OpenAI (safety card).
Anthropic: Claude Opus 5.5’s fast mode is now available to everyone. On the Claude developer platform, Opus 5.5’s fast mode has left research preview and the waitlist is gone: the same model responds up to 2.5 times faster, at $8 per million input tokens and $40 per million output tokens. Fast mode is also available in Microsoft Foundry (Azure) and is coming soon to Google Cloud; fast mode for Opus 5 and Opus 4.8 remains in preview. Source: Claude (X), Anthropic (release notes).
AI in science: $15 million for mathematicians, and CERN is building a physics foundation model with Mistral. The Simons Foundation, together with Anthropic and OpenAI, is launching $15 million in grants and resources so that mathematicians can evaluate results produced with AI, learn the new tools and develop shared rules for using AI. According to The Verge, OpenAI is contributing $5 million, a little over a week after it published hundreds of AI-written math manuscripts, which caused outrage in the field (we wrote about this earlier). CERN and France’s Mistral signed a letter of intent on long-term cooperation: together they would build the first particle physics foundation model, trained on billions of simulated collisions from the Large Hadron Collider (LHC); according to Reuters it will be called Flare, and no timeline or budget was given. OpenAI is also offering supported access to its life sciences model, GPT-Rosalind, to eligible US university labs under the name Rosalind Academic Plan. Source: Simons Foundation, The Verge, CERN, Reuters (CNA), OpenAI.
Image editing models: Krea 3 and Black Forest Labs’ open FLUX.3 Image Klein are here. Krea launched its Krea 3 image editing model in beta: according to the company, it completes 2K-resolution edits in under 8 seconds at about $0.06 per image, and changes only what is asked, for example within a marked area; Krea says its quality rivals Google’s Nano Banana Pro and OpenAI’s GPT Image 2.5. Germany’s Black Forest Labs released an open-weight model of roughly 6 billion parameters called FLUX.3 Image Klein: it generates from text and reference images, and the placement of parts of the image can be specified with bounding boxes; according to Hugging Face’s measurements, a 1024×1024 image takes about 2 seconds on a single H100 card. Source: Krea, Krea (X), Hugging Face.
What this means for you: The US FTC and international financial regulators are asking more and more specifically how companies oversee AI agents, so if you use one too, it is worth setting down now when human approval is required and logging what the agent does; our AI compliance consulting gives you something to hold on to. Meanwhile, open-weight models (Meta, Nvidia, StepFun) keep getting stronger and can even run on your own machines, so if you are planning to roll out an AI agent yourself, keeping your data inside the company is now a realistic option too. If you have a Facebook or Instagram business account and your own website, review Meta’s new terms before October 30.