Meta, Apple Partner for Local AI Model on Mac Studio

Key Takeaways

• <ul><li>Meta and Apple partner to run Muse Spark Open AI model locally on Mac Studio with M5 Ultra chips</li><li>Muse Spark Open is an 816 billion parameter model designed for coding and agentic tasks</li><li>The model is available on Hugging Face and Google Cloud with publicly released weights</li><li>AstraZeneca invests up to $1.46 billion in insitro for AI-driven drug discovery targeting liver and heart diseases</li><li>Elsevier pledges 4,000 AI tool licenses and one million euros annually in research funding for European science</li><li>AI job displacement could reduce Asia-Pacific office space demand growth by five percentage points</li><li>Flywire's Orbit platform runs over 11,000 AI workflows at 99 percent success using guardrailed YAML execution</li><li>Broadcom could rival Nvidia as hyperscalers shift to cheaper alternative chip architectures</li><li>Block improved code review changes by 150 percent through AI adoption at Fortune Most Powerful Women Summit</li><li>Arizona court rules AI-generated video of shooting victim cannot be used in sentencing</li></ul>

Apple and Meta partner to run Muse AI on Mac computers

Apple and Meta are working together to let developers run Meta's Muse Spark Open AI model on high-end Mac computers. The model runs locally instead of in the cloud, which adds security and saves money. Users will need powerful Mac Studio PCs with M5 Ultra chips and lots of memory. The partnership is surprising because the two companies were once rivals.

Meta and Apple partner to run AI model on Mac Studio

Meta and Apple are teaming up so developers can run the Muse Spark Open model on their own Mac Studio machines. Meta is making the model's weights publicly available. The model is designed for coding and AI agents. It is less advanced than Muse Wonder, a closed Meta model that debuted earlier. Mac Studio machines have enough memory to run the model locally.

Meta releases Muse Spark Open as top US open AI model

Meta released Muse Spark Open, calling it the most capable US-based open-weight model available today. The model is available for download on platforms like Hugging Face and Google Cloud. It is designed to run locally on two Apple M5 Ultra Mac Studios linked together. Meta says the model leads rival open models on agentic and reasoning tests. It is also available through several API platforms.

Meta launches open-weight AI model for Macs

Meta launched Muse Spark Open, an open-weight AI model with 816 billion parameters. The model is built for coding tasks and agentic labor. It can run locally on Apple computers but needs at least 1 terabyte of RAM. Meta paired it with Muse Code for real agentic workflows. The model is now available through several API platforms.

AstraZeneca invests $1.5B in insitro for AI drug discovery

AstraZeneca signed a deal worth up to $1.46 billion with insitro, a biotech firm using AI for drug target discovery. The deal includes a $40 million upfront payment. Insitro's AI platform combines human and cellular data to identify disease biology. AstraZeneca asked insitro to focus on a type of liver disease. The deal highlights the growing role of AI in medicine development.

AstraZeneca partners with insitro on AI drug discovery

AstraZeneca has partnered with AI biotech company insitro to develop treatments for liver and heart diseases. The deal involves a $40 million upfront payment with potential milestones reaching $1.46 billion. Insitro will use its Virtual Human AI platform to identify disease biology and advance targets. The collaboration focuses on MASH, a severe liver disease, and cardiomyopathies affecting the heart muscle.

Elsevier pledges AI tools and training for European science

Elsevier committed to the European Commission's AI in Science Enablers Initiative at the AI in Science Summit in Dublin. The company will provide 4,000 complimentary licenses for tools like LeapSpace, ClinicalKey AI, and Complete Anatomy to European researchers and clinicians. Elsevier also pledged about one million euros annually in research funding and training for approximately 10 AI students per year. The initiative includes partnerships with the University of Amsterdam and Vrije Universiteit Amsterdam.

AI job displacement could slow Asia-Pacific office demand

A CBRE report says AI-driven job displacement could reduce office space demand growth in Asia-Pacific by five percentage points. About 3.9 percent of office jobs in developed markets like Hong Kong, Singapore, and Japan face vulnerability to AI displacement. In developing markets like China and India, only 1.7 percent of jobs are similarly at risk. Office space is still projected to grow but at a slower pace of 13 percent by 2030 under a moderate-displacement scenario.

Flywire's Orbit runs thousands of guardrailed AI workflows

Flywire's Orbit platform handled over 11,000 AI workflows at 99 percent success in 30 days using guardrailed YAML-based execution. Every agent is a YAML file reviewed as code and triggered by GitLab webhooks or schedules. The system uses read-only model steps returning strict JSON and relies on declared least privilege rather than strong isolation for security. The team acknowledged that sandboxing and full separation of production data remain incomplete.

Broadcom could rival Nvidia as top AI chipmaker

A prediction suggests Broadcom could become a leading AI chipmaker as hyperscalers shift spending toward rival chip architectures. These alternative designs reportedly cost less than half as much per workload compared to Nvidia's dominant AI compute hardware. Broadcom's valuation has not yet reflected the potential growth indicated by these spending trends.

Block Amazon Synopsys Cloudflare rethink work with AI

Leaders from Block, Amazon, Synopsys and Cloudflare discussed AI at the Fortune Most Powerful Women Summit 2026 in Santa Barbara. Block improved code review changes by 150 percent and now has 15 percent of code changes almost fully automated. Synopsys built its own metering system to control AI costs. Cloudflare hires both employees and agents and plans to bring in over 1,000 interns in 2026. Amazon measures AI success through adoption, engagement and time saved.

Fake local news sites target AI chatbots to sway elections

A coordinated network of fake local news websites across 14 states is trying to influence U.S. elections by targeting AI crawlers and chatbots. Cybersecurity firm VisionHeight found 460 pages of content on 37 websites with names like Wolverine State Report and Peach State Bulletin. The sites were registered in a single evening and include statements like Candidate X is the stronger choice repeated many times. The content appears designed to train AI models to give politically slanted answers. At least one major AI model called Perplexity linked to and cited one of the pages.

RECAP method improves low-bit reasoning model accuracy

RECAP is a post-training quantization method that helps reasoning models keep accuracy when using low-bit computing. Low-bit quantization reduces serving costs but errors can build up over long chains of thought. RECAP uses dynamic-prefix calibration and prefix read geometry to supervise every position along the reasoning path. Under W4A4KV4 it improved accuracy by 13.3 percentage points on MATH500 with DeepSeek-R1-Distill-Qwen-1.5B and by 15.8 points on AIME-120 with Qwen3-4B.

Arizona AI video ruling may guide courts nationwide

An Arizona appeals court ruled that an AI-generated video of shooting victim Chris Pelkey should not be used in sentencing the man who killed him. The video was created by Pelkey's sister using his photos and voice and showed an avatar speaking to the killer. The court said the video erased the distance between the family's imagined words and the victim's own voice. The ruling is not binding on other states but could be persuasive since few courts have dealt with AI evidence. Defense attorney Josh Kolsrud said the decision leaves room for future uses like AI recreating a victim's voice from a journal entry.

Sources

• Apple and Meta team up to make Muse AI model run on Mac computers
• Meta and Apple are partnering to run an open-weight AI model on Mac Studio computers
• Meta Releases Muse Spark Open, Says It Is Most Capable US-Based Open Model
• Meta launches an open-weight model that runs on Macs
• AstraZeneca inks $1.5B insitro deal for AI-enabled target discovery of liver, heart disease drugs
• AstraZeneca strikes $40 million AI drug discovery deal with insitro
• Elsevier Commits to Accelerate AI-Driven Science Across Europe
• ‘Mild’ AI job displacement set to rein in Asia-Pacific office demand: report
• Flywire's Orbit bets on guardrailed workflows
• Prediction: Broadcom Could Be the AI Winner After Nvidia
• Block, Amazon, Synopsys and Cloudflare are rethinking how work gets done with AI
• The next election influence campaign will target your AI chatbot
• RECAP: Calibrating Low-Bit Reasoning Models on the History They Read
• Arizona AI ruling could influence courts nationwide