Local AI Hardware in Autumn 2026: What You Can Actually Run at Home

The number that decides which AI models a computer can run is memory. A large model has to fit in it, and most graphics cards top out well below what the biggest open models need. This autumn, several machines arrive with one big pool of memory shared between processor and graphics, up to 128GB or more. Here is what is on offer, what the makers say it can do, and what nobody has independently measured yet.

NVIDIA RTX Spark: October

NVIDIA says its first RTX Spark PCs arrive in October, as slim laptops and compact desktops from several manufacturers. The chip combines a 20-core Arm CPU with a Blackwell GPU of up to 6,144 cores and up to 128GB of unified memory, with up to one petaflop of FP4 AI performance, and it runs Windows 11 on Arm (NVIDIA).

Prices are not official. Press estimates for the top models sit around $3,000, citing a Morgan Stanley note, which makes them estimates rather than prices (PC Guide). Because it is Windows on Arm, check that the software you rely on works before you buy.

Apple's Mac Studio: M5 Max and M5 Ultra

Apple announced the new Mac Studio on August 25, and it started shipping on September 22. The M5 Max goes up to 128GB of memory from $2,499, and the M5 Ultra starts at $5,499. The 512GB M5 Ultra configuration arrives in late October and had no announced price at the time of writing (Kingy). That guide notes it had not measured either chip, and Apple's performance figures come from its own testing, so real-world model speeds are still to be seen.

AMD and others

This is not new this autumn, but it is the cheapest route. StorageReview's lab testing found that HP's Z2 Mini G1a, built on AMD's Ryzen AI Max+ PRO, ran a 120-billion-parameter model without a discrete graphics card (StorageReview). It shows that a small office box with unified memory can handle large models.

The software is getting easier

Hardware is only half of it. NVIDIA reports up to 1.9 times higher throughput in llama.cpp on a GeForce RTX 5090, and simpler local setup in the Hermes Agent, OpenClaw and Perplexity Portable Computer agent apps. It also released PAIR, a free open-source tool in beta that spreads AI requests across compatible computers on your home network. It runs on Windows, macOS and Linux and supports Apple M4 or newer (NVIDIA). These are NVIDIA's own figures.

NVIDIA also listed open models that run on this class of hardware, including Qwen3.8-27B and a 30-billion parameter coding model from Meta, and DeepSeek V4 Flash, a 284-billion-parameter model, on a pair of DGX Spark machines.

What to watch

• Independent benchmarks. Most speed claims so far come from the manufacturers.
• Pricing. RTX Spark prices and the 512GB Mac Studio price are not final.
• Memory bandwidth. It matters as much as capacity for how fast a model responds, and published figures for RTX Spark laptops vary between sources.
• Whether you need it. Smaller open models run on far less, so a machine like this suits people running large models or several agents at once.

What we don't know yet

We haven't tested any of these machines. The details above come from the manufacturers' announcements and early coverage, and they may change as products ship and independent reviews arrive.

More from Offgrid Studio

OpenAI's Dots: Always-On Agents That Work on a Computer of Their Own
How to Change OffgridStem's Language on iPhone (Add a Second Language First)