Apple has introduced two new desktop-focused chips that make its direction on artificial intelligence increasingly clear: more AI work is being pushed directly onto the device.
The company announced the M6 for the new Mac mini and the M5 Ultra for the new Mac Studio on August 25. Apple is positioning both chips around higher CPU and GPU performance, faster neural processing and much larger memory bandwidth for local AI workloads.
Apple M6 M5 Ultra and On-Device AI
The Apple M6 M5 Ultra launch shows two ends of the same strategy. M6 brings stronger local AI capability to a more mainstream desktop, while M5 Ultra targets workstation users who need much larger memory capacity and GPU throughput. Both chips make it easier for developers to consider workflows that keep more inference, experimentation, and private data processing on the Mac itself.
Apple’s official announcement emphasizes AI compute, unified memory, and local model workloads as central parts of the new hardware generation.
M6 Is Apple’s First 2nm Chip
According to Apple, M6 is its first chip built using a 2-nanometer process. It includes a 12-core CPU, a 12-core GPU and a Dual 16-core Neural Engine. Apple says the Neural Engine can deliver up to twice the peak compute of the previous generation.
The chip also includes Neural Accelerators inside the GPU cores and supports up to 32GB of unified memory with bandwidth of up to 170GB/s.
That combination is aimed at tasks such as coding assistants, image generation, local language models and agentic workflows that can run without sending every operation to a remote data center.
The bigger shift: Apple is not treating AI as a single “AI feature.” It is redesigning more parts of the chip — CPU, GPU, memory and Neural Engine — around workloads that increasingly mix traditional computing with machine learning.
M5 Ultra Is Built for Much Larger Local Models
The M5 Ultra targets a very different class of workload. Apple says the chip can be configured with up to a 36-core CPU, an 80-core GPU and 512GB of unified memory.
Memory bandwidth reaches 1.2TB/s, which Apple says is 50% higher than M3 Ultra. The company specifically highlighted the ability to keep very large datasets and large language models entirely in unified memory.
Apple says systems based on M5 Ultra can run LLMs with hundreds of billions of parameters locally, depending on model format and workload. That could be important for researchers, developers and companies that want to experiment with large models while keeping data on their own machines.
Why Local AI Is Becoming More Important
Cloud AI remains essential for the largest frontier models, but local inference has several advantages. It can reduce latency, work without a constant internet connection, limit recurring API costs and keep sensitive data on the user’s device.
The trade-off has traditionally been hardware. Running larger models locally requires substantial memory capacity and bandwidth. Apple’s unified-memory architecture has become one of the Mac’s most distinctive advantages for this type of workload because the CPU and GPU can access the same memory pool.
What Developers Get
Apple says its Core ML, Metal, Xcode and newer Core AI tools can access the additional compute in both chips. Developers can use Apple’s own Foundation Models or bring proprietary models to run and fine-tune locally.
This fits a broader industry trend. Windows PC makers, GPU vendors and Apple are all competing to make local AI a standard part of personal computing rather than a niche feature for enthusiasts.
This Is Also a Privacy and Cost Story
On-device AI does not automatically make every application private, because software can still send data to external services. But hardware capable of running more work locally gives developers the option to avoid cloud processing where it is not necessary.
For businesses, the economics may also become interesting. A powerful local workstation has a high upfront cost, but repeated inference on local hardware can reduce dependence on usage-based cloud APIs for certain workloads.
What to Watch Next
The hardware specifications are only part of the story. The next useful signal will be how quickly developers ship applications that can take advantage of the added memory and neural compute without compromising battery, thermals, or usability. Benchmarks should also distinguish between model loading, prompt processing, sustained generation speed, and total system power rather than reducing “AI performance” to one number.
For the infrastructure side of the same market shift, see our coverage of NVIDIA’s latest AI data-center growth and our biggest AI news overview.
The Bottom Line
M6 and M5 Ultra show Apple leaning harder into a future where Macs are expected to run meaningful AI workloads on-device. M6 brings faster neural processing to a more mainstream desktop, while M5 Ultra pushes memory capacity and GPU compute into workstation territory.
The most important question now is software. Hardware can make local AI possible, but developers still need applications that use the capability in ways that are genuinely faster, more private or more useful than simply calling a cloud model.
Source: Apple Newsroom, August 25, 2026.
Explore more: free online tools · AI tools directory · latest guides.
