Apple's new Mac Studio and Mac mini are built for local AI

No event, no keynote — just a surprise hardware drop with one very clear audience: people who run AI models on their desks.

Apple refreshed its entire desktop lineup around local AI inference today, debuting the M6 — its first 2nm chip — in a new Mac mini and the M5 Ultra, its first-ever quad-die design, in a new Mac Studio. The M6 brings a 12-core CPU and 12-core GPU with a Dual 16-core Neural Engine, and Apple claims the world's fastest single-threaded performance. The M5 Ultra goes further: it fuses two M5 Max dies with next-generation UltraFusion (over 4.4TB/s of inter-die bandwidth) into up to a 36-core CPU and 80-core GPU, with up to 512GB of unified memory moving at 1.2TB/s — 50 percent more bandwidth than M3 Ultra. That memory pool is the point: Apple says it can hold open-weight LLMs with hundreds of billions of parameters entirely on device. Pricing runs from $899 (Mac mini, M6) and $1,699 (M5 Pro) up to $2,499 (Mac Studio, M5 Max) and $5,499 (M5 Ultra); preorders opened today, shipping September 22.

Why this launch reads differently from a routine spec bump: ever since macOS enabled low-latency Thunderbolt 5 clustering for distributed MLX inference last December, developers have been daisy-chaining Mac minis and Studios to serve models far larger than any single consumer machine could hold — a budget-grade alternative to racks of specialized GPUs. This refresh, as Ars Technica observes, finally designs for that crowd rather than stumbling into it. Our take: with cloud token bills climbing and open models like Qwen and DeepSeek closing the capability gap, Apple is quietly claiming the "your desk is the datacenter" niche before anyone else thinks to contest it.

Researchers at Oasis Security disclosed a flaw in Nvidia's NemoClaw agent toolkit that let a single visit to a malicious website hijack the local model behind a developer's AI agent. NemoClaw launches its Ollama model server bound to every network interface with no authentication, so DNS rebinding let an attacker's page reach it through the victim's own browser — then rewrite the model's prompt template so a hidden instruction rode along beneath every message, invisible to guardrails and untouched by clearing the chat. Oasis reported it to Nvidia's security team before publishing, but the lesson outlives the patch: sandboxing the agent means little when the plumbing underneath is reachable from any tab.

Keenable, founded by former Yandex search chief Andrey Styskin, exited stealth with a $26 million seed round led by Accel to build a web search index designed for AI agents. The company claims more than 100 billion documents indexed, already used in production by several AI labs and inference providers — well timed, given Google and Microsoft have been retiring the search APIs much of the agent ecosystem leaned on. Styskin says the dream is becoming the next Google for AI agents, beating the incumbent on machine-driven queries where human-oriented result pages fall short.

What to watch: whether the top-end 512GB M5 Ultra configuration — which slips to late October — becomes the de facto entry point for running frontier-scale open models outside the cloud.

Would you run your next coding agent on a desk-sized Mac cluster instead of renting tokens? Tell us in the comments.

Brain organoids classified Braille at 83% — in a trio, not alone

Two stories today: human brain tissue on a chip decoded a tactile signal, and the data-centre company behind some of the largest AI contracts in Europe turns out to be six months behind on its flagship site. Human forebrain organoids grown on a microelectrode array classified Braille letters with 61 percent accuracy on their own and 83 percent when the responses of three organoids were combined, in a peer-reviewed study from the University of Bristol. The team took event-based tactile readings

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Microsoft finished pre-training its biggest in-house model yet — then used it to make a point about the company it bankrolls. Microsoft AI completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters, and says the recipe is 4.2–12.3x more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base. The run itself is the number worth keeping: 16,000 Nvidia GB200 GPUs for 38 days across August and September, at 21% average model

Amazon's new Bee devices listen all day — for $11.99 a month

Amazon turned its always-listening Bee line into a two-device family on Thursday, and NVIDIA opened a 2.8-trillion-token Indic corpus recipe. The connecting thread: both are bets that the next AI money is in memory — yours, or your model's. Amazon put two new always-listening devices on sale, and both now lead to a subscription. Bee Frontier, the next-generation wristband that also clips to clothing, ships now at $74.99 in Onyx, Lavender and Sage — up from the $50 predecessor — with seven-day

Perplexity open-sources its agent sandbox after 118 failed escapes

Perplexity handed its agent containment layer to the open-source world, and Washington's AI task force stopped reading briefings and started working. Perplexity open-sourced SPACE on Thursday — the sandbox it built to keep its own AI agents from getting out — and says nine frontier models failed to break out of it across 118 attempts. Each agent runs inside a small virtual machine with its own operating system, with credentials and network access managed outside the sandbox so an agent can re

Read more

Brain organoids classified Braille at 83% — in a trio, not alone

Suleyman says Microsoft's base model beat OpenAI's on a quarter compute

Amazon's new Bee devices listen all day — for $11.99 a month

Perplexity open-sources its agent sandbox after 118 failed escapes