Local Artificial Intelligence (AI) is becoming easier to use on Windows PCs and Macs. Microsoft is adding a local model to PowerToys, Apple is positioning new desktops more explicitly for local AI, and new measurement methods aim to show how compact models actually perform on real devices. This matters if you want to process text or other data with less dependence on the cloud while balancing speed, cost, and privacy.
Why local AI matters
With local AI, the model runs on your own device rather than entirely on a remote server. Individual steps can therefore work without a continuous connection to a cloud service. Whether data truly stays on the device still depends on the whole workflow: If an application uses additional online services or interfaces, information may continue to leave your computer.
The most immediate benefit is direct availability. Local processing can also avoid usage fees for every piece of text processed, provided that the relevant step runs exclusively on your hardware. You pay indirectly through the computer, its electricity use, and storage requirements, while more capable models generally place greater demands on those resources.
An everyday example is copying text from a screenshot. The updated Advanced Paste feature in Windows PowerToys can extract text from an image and insert it directly. A workplace example comes from Mac users who connect several machines to run language models that are too large for one ordinary consumer device.
Local AI does not automatically mean better AI. Small or heavily compressed models may lose quality in reasoning, mathematics, and code generation. A quick response is not proof of accuracy either: A model can answer promptly and still be factually wrong.
What Windows now makes easier
PowerToys is a Microsoft-provided collection of additional tools for Windows 11. It adds features that let you customize the operating system and simplify repetitive actions. Its latest update gives Advanced Paste a more accessible local AI option.
The feature could already use AI to transform copied material while you inserted it. Cloud-based AI required users to configure application programming interface keys and endpoints; an application programming interface connects software services, while an endpoint is the specific address used to reach one. The local option removes that setup step.
Instead, you download Phi Silica once, a local AI model supplied by Microsoft. Supported processing can then take place on the Windows 11 PC. The report provides no model price, detailed hardware requirements, supported-language list, or comparative quality results, so the announcement does not yet show how consistently the feature performs across different computers.
Step 1: Set up PowerToys and the model
- Install or update Microsoft PowerToys on Windows 11.
- Open the Advanced Paste settings and download the local Phi Silica model.
- Check whether the application identifies your intended operation as a local feature.
Step 2: Test it with limited content
- Start by copying a short passage or creating an image containing clearly legible text.
- Use Advanced Paste to transform the content or extract and insert the text from the image.
- Compare the result with the original before using the feature for important documents.
This is a practical entry point for beginners because you do not need to supply your own cloud-service keys. The local label should not encourage blind trust, though: Check the settings and output when dealing with confidential material. The report does not say whether every feature can operate entirely offline.
What the new Macs offer for local AI
Apple is positioning the M6 Mac mini and M5 Ultra Mac Studio more directly for local AI inference. Inference is the stage in which an existing model processes an input and produces an answer. According to Ars Technica’s account of the new Mac hardware, these workloads benefit from Apple’s unified memory, which is shared by central and graphics processing units.
The M6 is Apple’s first Mac chip manufactured with a 2-nanometer process and has 12 central processing cores: two cores Apple calls “super cores,” four performance cores, and six efficiency cores. It also has 12 graphics cores, each with its own neural accelerator. The comparison of the M6 and M5 Ultra architectures describes two different approaches: The M6 targets efficiency in the Mac mini, while the four-die M5 Ultra is intended for more complex workloads in the Mac Studio.
Apple says the M6 Mac mini delivers up to four times the AI performance of the previous generation and processes memory data twice as quickly. The supplied sources do not independently verify those manufacturer claims. Ars Technica also offers a more restrained assessment, describing the computers as specification upgrades without major new device features.
Pricing further qualifies the idea of an easy entry point. According to t3n, the new M6 Mac mini starts at more than €1,000, while the M4 model originally cost €699. None of the sources provides Swiss pricing or availability, so the euro figure should not be treated as a local retail price for Switzerland.
Since macOS 26.2, suitably equipped Macs have been able to communicate with low latency over Thunderbolt 5, a fast wired connection, for distributed AI inference. MLX, an open machine-learning framework for Apple chips, uses their unified memory. Hobbyists and professionals have connected multiple Mac minis or Mac Studios to run larger local language models; that can provide substantial capacity, but it is hardly the inexpensive default for occasional text work.
Why models and measurement methods belong together
Model size alone tells you little about how useful local AI will be on your device. Model cards often measure quality with powerful server hardware and high numerical precision. According to the announcement of the open Pipette benchmarking platform, those results rarely predict how the same model will behave on a phone or another edge device.
Pipette is designed to measure the model, quantization, runtime software, and hardware together. Quantization stores model values with reduced precision to lower memory and computing requirements. Liquid AI developed the open platform with Artificial Analysis, which the report identifies as an independent validator of the methodology, although the supplied summary includes no actual benchmark results.
Another approach attempts to repair the quality loss that commonly follows model compression. The Multiverse Computing team describes a method called Quantization-Aware Healing, a recovery process that explicitly accounts for quantization. A model compressed to 60 billion parameters and reduced to 4-bit precision beat its own higher-precision version in seven of nine tests, according to the published account of the compression work.
That result challenges the common expectation that a 4-bit model must always perform worse than the more precise version from which it came. It applies only to the described model and training setup, however, and the supplied sources do not independently confirm it. For you, the practical lesson is not that compression always improves quality, but that a model name and file size are insufficient for a fair comparison.
At the other end of the spectrum, Perplexity’s “Portable Computer” combines local models, a control environment, connectors to other services, and operating-system-enforced isolation on an Nvidia DGX Spark. This isolated environment, also called a sandbox, limits what running software can access. The brief Portable Computer announcement claims zero per-token cost for local steps, but gives neither a purchase price nor independent performance measurements.
Pros and Cons of local AI
Pros:
- Less cloud dependence – Supported tasks can run without sending a continuing request to a remote AI service.
- No usage-based fees – Fully local steps carry no charge per piece of processed text, although hardware and electricity still cost money.
- Shorter data paths – Entirely local processing does not need to send material to a server and back.
- More choice – Windows utilities, compact Macs, and specialized systems now address different levels of demand.
Cons:
- High entry costs – Capable hardware can be expensive, as the Mac mini’s price increase illustrates.
- Variable quality – Compression saves memory but can weaken reasoning, mathematics, and code generation.
- Inconsistent comparisons – Manufacturer figures and server tests do not reliably represent performance on your device.
- No automatic privacy – Online connections, extensions, or incorrect settings may transmit data even when the core model is local.
What this means for you
If you are just starting, choose one clearly limited task on hardware you already own. On Windows, extracting text from an image with PowerToys is one practical option. Compare the output with the source, observe response time and memory use, and avoid confidential documents until the application’s data flow is clear.
If you already use local models, you can get more meaningful results by looking beyond model size. Consider quantization, runtime software, available memory, and the exact hardware. Reproducible tests of the kind Pipette aims to provide are more informative than one speed figure measured under server conditions.
For Switzerland, the sources provide no specific information about prices in Swiss francs, product availability, supported national languages, or compliance with local data protection rules. Local processing can reduce how much data is transmitted, but it does not prove that an application meets Swiss privacy requirements. For school materials, customer information, or internal documents, the decisive issue remains whether the entire workflow is genuinely local.
Local AI is moving from a specialist project toward an accessible computer feature, although performance and prices vary widely. PowerToys lowers the initial barrier, while new Macs and specialized systems can accommodate larger models; improved benchmarks may make those options easier to compare. The unresolved risk is whether vendor claims reflect everyday use and whether local tools can match their cloud counterparts in quality, language coverage, and privacy.
Sources
- Windows 11: Update für Powertoys bringt neue Features und lokale KI – t3n, 2026-08-25
- Liquid AI Open-Sources Pipette: A Reproducible Benchmarking Suite – MarkTechPost, 2026-08-25
- Perplexity Ships Portable Computer on NVIDIA DGX Spark – MarkTechPost, 2026-08-25
- Apple M6 und M5 Ultra: Die neuen Desktop-Chips im Architektur-Vergleich – t3n, 2026-08-25
- Mac Mini mit M6 vorgestellt: Das Gehäuse bleibt klein – der Preis nicht – t3n, 2026-08-25
- Apple’s new desktop computers are designed specifically for local AI development – Ars Technica, 2026-08-25
- Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original – Hugging Face, 2026-08-25
