Apple and Meta Cooperate Against Chinese Open-Source Models

Apple and Meta End Dispute: Meta’s 816-Billion-Parameter Model Runs on Mac Studios

Meta released Muse Spark Open on Thursday, publicly collaborating with Apple for the first time. The model has 816 billion parameters, of which 42 billion become active per token via a Mixture of Experts architecture, along with a context window of one million tokens. According to Meta, it is the most powerful open-weight model developed in the US and is designed for long task chains, multi-step reasoning, and software development; it processes text, images, video, and audio. On the JobBench benchmark, the provider reports a score of 61.8, putting it ahead of GLM 5.3. The weights are available for free download under a license with limited commercial use.

The technical core of the collaboration is running the model locally. Two to four coupled Mac Studios with Apple’s M5-Ultra chip and the MLX framework hold the model in memory instead of running it in a data center. Apple points to its Unified Memory Architecture, where the integrated GPU addresses up to 512 gigabytes per machine, with a memory bandwidth of 1.2 terabytes per second. Meta describes the result as two computers that behave like a single device with one terabyte of shared memory. In parallel, Apple released new benchmarks for its chips and technical documentation on how MLX distributes open models across multiple machines.

The setup is not cheap. A Mac Studio with M5 Ultra and 256 gigabytes of memory currently costs at least $10,799, the 512-gigabyte version is expected to be released later this month, and users need multiple devices. The weights are available via Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models, with hosted access offered by Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, among others. In August, Meta had already released Muse Glimmer, a smaller model for regular notebooks.

The rapprochement between the two companies is remarkable against the backdrop of a long-standing conflict. In 2021, Apple introduced privacy features that restricted ad industry targeting; Meta estimated the expected revenue loss for 2022 at ten billion dollars. Meta’s AI chief Alexandr Wang said both sides have been working together for several months, naming Daniel Gross, a former Apple manager who is now at Meta, as a central figure in the mediation. Mark Zuckerberg publicly thanked the new Apple CEO John Ternus, who returned the greetings. Neither company would say whether this will turn into a deeper partnership. In parallel, Nvidia, Reflection AI, and Mistral have released their own open models, while Washington is debating whether freely available Chinese models are undercutting American providers. → cnbc, axios, MacRumors, The Verge, techstrong

Synthszr Take: A 41 percent download share for Chinese models—that’s the number that explains this release, not some breakthrough in Meta’s research. Zuckerberg already demonstrated with Llama how to dry up a market that others want to monetize, and the pattern is now repeating itself in reverse: this time, Meta is defending terrain that Qwen, DeepSeek, and Kimi already occupy. The license is telling: above $20 million in revenue, you first have to negotiate with Meta and are supposed to display 'Built with Muse Spark Open' everywhere. Meta is selling marketing reach and calling it openness. The deficit in terminal coding (76.4 versus 85.0 for Kimi K3) has been an unresolved issue since April, and the desktop demo with two Mac Studios for the price of a small car is a stunt for journalists, not a procurement option. This market will be decided where developers download and fine-tune weights, and right now, Hangzhou is leading, not Menlo Park.

Nvidia’s Open Nemotron Model Partially Beats Claude Opus 5

Nvidia has introduced Nemotron 3.5 Super VL, an open model with 120 billion parameters that processes text and visual content such as PDFs, diagrams, tables, and web pages in a single pass. Additionally, there are two smaller models for locally running agents: Bolt 74B-A9B for coding and research tasks, and Bolt E4B with 4 GB, which is set to be released on October 23. According to the provider, companies can adapt Super VL to their own documents, tools, and policies via post-training and then operate it where their data residency requirements allow. A graphic published by Nvidia shows five specialized tasks where the adapted variants perform between 9 and 44 percent better than the base model; for medical insurance claims, the accuracy increases from around 28 to 72 percent, compared to about 38 percent for Claude Opus 5, and for attack resistance, from 35 to 76 percent, compared to around 66 percent for GPT-6.1 Sol. Nvidia provides neither task definitions, sample sizes, nor test protocols, and according to VentureBeat, these are also different adapted versions for different specialized tasks. → VentureBeat

Synthszr Take: The base model has an accuracy of around 28 percent for insurance claims, which is useless in any claims department in the world. The 72 percent that follows primarily describes the work that comes before: annotated files from one’s own records and internal review rules, translated into test cases against which the result can even be measured. The partner list shows where this pays off: in credit agreement analysis, due diligence reviews, and supplier credits—processes with consistently similar document types and written decision rules.

Anthropic’s Growth Is Slowing: Investors Panic

The Financial Times has obtained Anthropic’s quarterly figures from a private investor meeting. According to the figures, the company generated over $16 billion in revenue in a single quarter, with 46 percent growth compared to the previous quarter and a reported run rate of around $74 billion. The crux is in the comparison: in the previous quarter, growth was still at 145 percent, so the curve is visibly flattening. The Algorithmic Bridge interprets this as a transition from an exponential trajectory to a Sigmoid curve and considers the internal target of a $100 billion run rate to be barely achievable, unless the fourth quarter surprisingly picks up again. At the same time, according to the report, there are infrastructure commitments of $518 billion over the next decade, of which around 80 percent are non-cancelable. → The Algorithmic Bridge

Synthszr Take: $16 billion in revenue in three months, and the market reads it as a warning sign. This is what happens when a company makes its own growth curve the foundation of its business: a drop from 145 to 46 percent would be a dream quarter elsewhere, but here it’s a crack in the promise on which $518 billion in infrastructure commitments rest, four-fifths of which are non-cancelable. Exponential growth as an expected value has a nasty side effect: from that point on, every second derivative becomes a headline, and success only counts as a deviation from the prospectus.

CoreWeave Plummets by 15 Percent

Shares of AI data center provider CoreWeave fell by 15 percent on Wednesday after a chart showing the company’s rising financing costs circulated on X. Venture capitalist Chamath Palihapitiya had shown that the yield on a $1.25 billion CoreWeave bond maturing in 2032 temporarily touched 15 percent and closed at a new high of 14.5 percent, raising the question of whether the AI build-out is creating a bubble in private credit. Elon Musk responded with one word: 'Yikes'. According to Bloomberg, CEO Mike Intrator wrote in an internal memo that the market misunderstands the situation and is acting 'irrationally,' pointing to weakness and delays in competitors' IPOs and debt financing that were coloring the narrative about CoreWeave. CoreWeave rents Nvidia chips to AI companies and finances the hardware predominantly with debt: interest expenses rose to $640 million in the second quarter from $267 million a year earlier, the quarterly loss was $626 million, and revenue more than doubled to $2.58 billion. Total debt reached around $52 billion at the end of June, up from about $15 billion a year earlier. → Yahoo Finance

Synthszr Take: Intrator is right on one point: the slide was triggered by a chart on X that the world’s richest man commented on with a single word, not by any newly released numbers from the company. That competitors' postponed IPOs would color their own narrative is observable behavior in credit markets, which tend to hold entire industries collectively liable. But the memo doesn’t answer the question of why creditors are now demanding 14.5 percent for bonds maturing in 2032, while the quarterly interest burden has more than doubled from $267 million to $640 million within a year.

Developer Rebuilds Photoshop and PowerPoint Using AI

Programmer Brandon Thomas from Atlanta has created and released free replicas of common Adobe programs using AI systems: Photoshop became “Photocraft,” Premiere became “Filmcraft,” for a total of seven applications initially. According to his own account, the trigger was frustration with Adobe’s subscription model. The clones offer fewer features than the originals but are designed for AI use from the ground up and can be controlled remotely via ChatGPT or Claude. Thomas has since rebuilt more programs through his company Artcraft, including “Deckcraft” as a PowerPoint replacement and “Wordcraft” for Microsoft Word. Legally, the case may be difficult to tackle because Artcraft apparently does not copy source code but imitates functions—a practice known as Reverse Engineering that has long been common in fields like mechanical engineering. → Handelsblatt KI-Briefing

Synthszr Take: A single developer from Atlanta has rebuilt seven Adobe programs, and Adobe’s strongest line of defense, copyright on the code, probably doesn’t even apply here. What Adobe has been selling for years was essentially a list of features, and today that can be reproduced in days. What remains is what Thomas can’t deliver in a weekend: maintenance across ten years of file formats and system updates, plus liability to an agency that earns its money with it. Asha Sharma at Xbox understood this faster than the subscription managers in the creative industry: she measures success by whether more people create things themselves, and uses Xbox Labs to build a feedback loop from that, where others would call their lawyers.

Perplexity Is Giving Away Its Agent Sandbox SPACE as Open Source

Perplexity is making SPACE, its own sandbox for securing AI agents, freely available as an open-source project. This was reported by Semafor, to whom the company confirmed the move on Thursday. SPACE is an isolated runtime environment where models and agents can be executed without being able to access the rest of the system. Perplexity names CoreWeave, Akamai, and Crusoe as partners who want to offer the tool to their own customers. According to the company, nine frontier models were instructed to break out of the environment in 118 attempts: none succeeded, whereas the models did escape from other, non-open-source sandboxes. → Semafor

Synthszr Take: There’s no business for Perplexity in sandboxes alone, because container isolation is infrastructure, and only those who write support contracts afterward can sell infrastructure. Perplexity sells answers and agent runs, and for that, a widely adopted breakout protection is cheaper than a single public incident that could suffocate the entire agent market. Added to this is the sober operational reason the company itself cites: strict isolation consumes computing time, and if CoreWeave, Akamai, and Crusoe deploy the code on their systems, external teams will help optimize Perplexity’s own cost base.

OpenAI Is Rolling Out Ads in ChatGPT: Free Users Are Targeted via Their Prompts

OpenAI has confirmed that ads are being gradually rolled out to users of the Free and Go tiers in selected regions; paying subscribers in Plus, Pro, Business, Enterprise, and Edu will not see ads. According to the provider’s help documentation, the ad selection is influenced by the context and intent of the current conversation, topic descriptions provided by advertisers, the ad content itself, and, if personalization is enabled, other signals from the user’s overall ChatGPT activity. Advertisers provide so-called “context hints,” which are descriptions of conversations, topics, or keywords where their product might be relevant; according to OpenAI, these hints guide the matching but do not guarantee placement in a specific conversation. An investigation by 404 Media reports that advertisers can also upload their own customer and prospect lists to create Custom Audiences. According to the documentation, this supports raw email addresses, phone numbers, their SHA-256-hashes as well as Google Advertising IDs, with inclusion targeting requiring a minimum of 25,000 matched users. → Gadget Review

Synthszr Take: The price of the free tier is now stated in the help documentation, and it is: intent. Google tied the invisible bond between intent and ad through a search box with three keywords; ChatGPT receives the same intent delivered in entire paragraphs, including symptom descriptions, salary questions, and draft job applications. Turning off personalization looks like control, but it still allows the current thread to be used as a signal: there is no real opt-out for Free and Go.

Reuters: The Order to Throttle Starshield for Ukraine Came from the White House

The instruction to slow down the SpaceX-operated satellite service Starshield for the Ukrainian military came from the White House, according to Reuters. Two US officials and a Western representative told the agency that the Pentagon leadership had received such a message; the anonymous sources could not name who in the White House made the decision or who implemented it. Ukrainian authorities reported this week that military Starlink receivers on fast-moving objects were refusing service, which affected the on-board guidance of drones. A representative from the Ukrainian Ministry of Defense said on Thursday that the receivers were working again while in motion, but only up to a speed of about 160 km/h. When asked, a senior US official stated that such an order was never issued by the president; the White House did not comment on other government employees. → www.reuters.com

Synthszr Take: Throttling to 160 km/h is a political decision, and so far, no one wants to have signed off on it. Two US officials point to the White House, a senior representative says the president didn’t demand anything, and the White House remains silent about any other conceivable source of the order. Musk has cleanly handed back responsibility: ownership and control lie with the US government, so either the military or the president decides—which is formally correct but practically worthless as long as there is no name behind it.

Tech and Telecom Companies Pay $183 Million for Trump’s Ballroom

The Washington Post reports it has obtained a complete donor list for the ballroom planned by Donald Trump at the White House; over $183 million of it comes from the technology, telecom, and crypto industries. According to the report, the largest individual donors are Nvidia and Apple with $25 million each, as well as Dish co-founder Charles Ergen, also with $25 million. Microsoft, Meta, Palantir, Google, Amazon, and Cisco are also on the list with multi-million dollar contributions, along with BlackRock, Lockheed Martin, and Coinbase. The payments are channeled through a non-profit organization directly controlled by the president and are not subject to disclosure requirements or campaign finance laws. About $77 million of the total amount cannot be attributed to any donor. → Engadget

Synthszr Take: $25 million from Nvidia, $25 million from Apple: compared to an exemption from semiconductor tariffs or Apple’s $600 billion investment commitment, that’s petty cash with an exceptionally good return. What is being paid for here is access in an administration that treats tariffs, spectrum issues, and export licenses as bargaining chips, not a building. For the corporations, this is a rational calculation; for the market, it’s toxic because competitive advantages are thereby awarded in Washington instead of in the product.

Accenture Caps Claude Budget at $29 Per Employee

Since October 1, the consulting firm Accenture has only been funding Claude Enterprise for its employees at a rate of $29 per person per month. This is according to an internal announcement obtained by 404 Media. Billing is usage-based: anyone who exceeds the $29 must find a project or team to cover the difference via a Work Breakdown Structure. Employees on the premium plan have a weekly token budget of $250, but many are on the general plan. In internal messages, individual employees report monthly costs between $500 and $1,000, one of them for building an AI accelerator for several Accenture clients; another writes that his Claude access has been completely revoked. → 404 Media

Synthszr Take: For someone who has been running up a $1,000 bill, $29 a month is about two working days of Claude Code. The reaction of those affected is procurement Tetris: the upgrade page is blocked, so they look for a project number (WBS) to take the bill, and when in doubt, it ends up with the client. This hardly changes usage behavior; only a portion of the most productive people’s working time shifts from building to justifying.