Nvidia’s Nemotron 3 Super, Perplexity’s Mac Mini Agent & Microsoft Backs Anthropic

NVIDIA Launches 120B Nemotron 3 Super Model 🧠

NVIDIA just built the perfect brain for AI agents. Moving beyond its hardware dominance, Nvidia has released Nemotron 3 Super, a highly advanced, open-weight 120-billion-parameter model designed specifically to solve the biggest bottlenecks in autonomous agent workflows.

Here is why this model is a massive leap for agentic AI:

• Solving “Context Explosion”: Multi-agent systems pass massive amounts of data back and forth, often causing the AI to “forget” its original goal. Nemotron 3 Super features a 1-million-token context window, allowing it to swallow entire codebases or thousands of financial reports without losing track of its objective.
• Hybrid MoE Architecture: The model uses a highly efficient “Latent Mixture-of-Experts” design. Despite having 120 billion parameters, it only activates 12 billion at inference.
• The Speed: When optimized on Nvidia’s Blackwell chips, the model delivers 5x higher throughput and predicts multiple tokens simultaneously, making it incredibly fast and cheap to run.
• The Adoption: The model is completely open-weight. Major platforms like Perplexity and enterprise giants like Palantir and Salesforce are already integrating it to run complex background workflows in cybersecurity and software development.

Why it matters: Nvidia realized that “Agentic AI” is too expensive if developers are forced to use massive frontier models for every tiny sub-task. Nemotron 3 Super is designed to be the ultra-fast, highly efficient “middle management” layer of the AI economy, solidifying Nvidia’s role as the primary software provider for the agent revolution.

UrviumAI Take: This is how Nvidia sells more Blackwell chips. Pay attention to the “NVFP4 precision” detail. Nvidia optimized this model specifically to run 4x faster on its newest Blackwell architecture. By releasing an incredible open-source model that works best on their own proprietary hardware, Nvidia is brilliant at driving demand for their multi-million dollar server racks.

Perplexity Turns Mac Mini into a 24/7 AI Agent 🖥️

Your next home appliance might just be a brain. Capitalizing on the viral craze of local AI agents, Perplexity has unveiled “Personal Computer,” a localized version of its multi-model orchestration system designed specifically to run 24/7 on a dedicated Mac mini.

Here is why Perplexity is trying to own your desktop:

• The Hardware Play: Instead of running agents purely in the cloud, the Personal Computer system turns a physical Mac mini into a persistent local agent. It has full access to local files, applications, and sessions.
• The “Safe” OpenClaw: Perplexity is explicitly positioning this as the enterprise-grade, secure alternative to the viral OpenClaw project (which famously went rogue and deleted a Meta executive’s emails). It features strict safeguards like tracked activity logs, mandatory sign-offs for sensitive tasks, and a physical ‘kill switch.’
• Remote Management: You don’t need to be sitting at the Mac mini to use it. Users can manage the local agent securely from anywhere, treating the machine like a remote digital employee.
• The Rollout: The system is currently rolling out to Perplexity Max subscribers via a waitlist, alongside a broader enterprise release that taps into 20 different AI models and over 400 app integrations.

Why it matters: Apple has been heavily criticized for lagging in the AI software race, but the Mac mini is quietly becoming the undisputed hardware champion of the agentic era. Because it is cheap, powerful, and easy to leave plugged in 24/7, AI startups like Perplexity are using Apple’s hardware as the Trojan horse to put autonomous, localized AI agents into millions of homes and offices.

UrviumAI Take: This is the bridge between software and hardware. If you want to experiment with local AI agents but are terrified of them deleting your files, this is the safest entry point. Buying a $599 Mac mini and dedicating it entirely to Perplexity creates a physical “sandbox.” If the bot goes crazy, it only breaks the mini, not your primary work laptop.

Microsoft Backs Anthropic Against Pentagon ⚖️

Big Tech is uniting against the Pentagon. In a massive show of industry solidarity, Microsoft has officially thrown its legal weight behind Anthropic in its escalating federal lawsuit against the US Department of Defense.

Here is why Microsoft is stepping into the crossfire:

• The Amicus Brief: Microsoft filed a motion in a San Francisco federal court urging the judge to issue a temporary restraining order against the Pentagon’s decision to label Anthropic a “supply chain risk.”
• The Argument: Microsoft warned that the Pentagon’s abrupt ban on Claude is reckless. They argued that forcing defense contractors to immediately rip out Anthropic’s technology would cause massive operational disruption, potentially “hampering U.S. warfighters at a critical point in time.”
• The Underlying Conflict: Anthropic was blacklisted for refusing to drop its ethical guardrails (specifically bans on mass surveillance and autonomous weapons). The DoD wants “all lawful uses,” and threatened to banish Anthropic if they didn’t comply.
• The Stakes: Microsoft—despite being a primary backer of OpenAI—invested up to $5 billion in Anthropic last November. Beyond protecting its investment, Microsoft is trying to protect the concept of corporate cloud sovereignty from federal overreach.

Why it matters: This isn’t just about Anthropic; it’s about the entire tech sector pushing back against government coercion. Microsoft is signaling to the Pentagon that you cannot simply rip out the foundational AI software running modern logistics because of an ideological disagreement without breaking the military’s own supply chain in the process.

UrviumAI Take: This is a brilliant legal maneuver by Microsoft. Notice the framing. Microsoft didn’t argue about “AI Ethics” or “Free Speech” like Anthropic did. They argued about “Military Readiness.” By telling a federal judge that the Pentagon’s ban is actually hurting the troops by breaking their software, Microsoft gave the court a massive national security justification to pause the blacklist.

Last AI News: Meta Acquires Moltbook, Nvidia’s NemoClaw, & Thinking Machines’ 1GW Deal

Other AI News Today:

• Cloudflare introduced a new /crawl API endpoint that allows developers to scrape entire websites with a single call, marking a massive pivot for the cybersecurity giant into the AI data business.
• Elon Musk announced the merger of xAI’s Macrohard project with Tesla’s Digital Optimus to create an AI system designed to emulate the functions of entire companies.
• Anthropic launched the Anthropic Institute, a ~30-person think tank designed to study AI’s societal disruption and engage with displaced workforces.
• Replit raised $400M at a $9B valuation and launched Agent 4, a powerful coding assistant that builds software 10x faster using parallel execution and an infinite design canvas.
• Amazon enforced a 90-day code safety reset requiring dual sign-offs after its AI coding tool Q was linked to multiple costly e-commerce outages.

Jigar Chaudhary

Jigar Chaudhary is the Editor-in-Chief at UrviumAI, where he oversees coverage of artificial intelligence news, tools, and in-depth studies. With over 5 years of experience analyzing AI and robotics, he focuses on maintaining high editorial standards, accurate reporting, and clear explanations to help readers understand how AI is shaping the future.