AI · Explainer
Strix Halo for local AI, explained
Updated 19 July 2026 · about a 7 minute read
Strix Halo is the chip that got the local-AI crowd properly excited, and for once the excitement's earned. It's AMD's Ryzen AI Max+ 395 - a laptop-class processor with a big graphics core and, the part that matters, up to 128GB of fast memory the GPU can reach into directly. So a mini-PC the size of a hardback holds a 70-billion-parameter model that would choke a £1,900 graphics card. The catch, as ever, is speed.
Here's what it is, what it'll actually run, which little box to buy, and where a GPU or a Mac still beats it.
What Strix Halo actually is
It's one chip doing three jobs. A Zen 5 processor, a big Radeon 8060S graphics core built in, and a separate NPU for AI maths - all on the one package AMD calls the Ryzen AI Max+ 395, codename Strix Halo. On its own that's a nice laptop part. What makes it interesting for us is the memory: up to 128GB of LPDDR5X-8000 soldered on a 256-bit bus, and the graphics core can be handed as much as 112GB of it for models.
It's the same idea behind a Mac Studio or a DGX Spark: one big pool of memory, shared, instead of a separate slab of VRAM on a card. It's why a machine you could post through a letterbox can hold a model that a 32GB RTX 5090 simply can't. The difference is what you pay and how fast it runs, and that's where the choices start.
What it'll run, and how fast
Anything that fits in the memory, which is almost everything short of the giant frontier models. The limiter isn't capacity, it's bandwidth. That 128GB runs at around 256 GB/s on paper and closer to 210 in the real world, which is a fraction of a graphics card's. So the machine fits a huge model that a GPU can't touch, then reads it out at a steady walking pace rather than a sprint.
Jeff Geerling's benchmarks put real numbers on it, and they tell the story clearly - the models that spread their work out (the mixture-of-experts sort) fly, while dense models drag once they get big:
| Model | Single-stream speed |
|---|---|
| gpt-oss-20B (MoE) | ~45 tok/s |
| Qwen 14B (dense) | ~23 tok/s |
| gpt-oss-120B (MoE) | ~33 tok/s |
| Llama 70B Q4 (dense) | ~5 tok/s |
Figures from Jeff Geerling's Strix Halo testing, single node, at roughly 100 to 140 watts. Your numbers will shift with quant, context and backend.
So the sweet spot is a big mixture-of-experts model, or a mid-sized dense one you're happy to read at conversational pace. A 70B dense model at about 5 tokens a second is usable for a considered back-and-forth, not for hammering. If you need fast, you need a graphics card, and you accept its smaller memory. That trade is the whole decision.
Which box to buy
Strix Halo is a chip, not a product, so you buy it inside a mini-PC. They're near-identical under the lid - same 128GB, same limits - so it comes down to price, ports and cooling. The three I'd look at:
Value
GMKtec EVO-X2
The cheapest way onto 128GB. Same chip, keenest price.
Read the review →
Flexible
Framework Desktop
Standard-parts x86 box from a repair-friendly brand.
Read the review →
Premium
Beelink GTR9 Pro
Dual 10GbE, dual USB4, better sustained cooling - at a price.
Read the review →
For most people the GMKtec is the sensible buy - it's the same silicon and memory for the least outlay. Pay up for the Framework if you value a repair-friendly brand and standard parts, or the Beelink if its 10-gigabit networking and sturdier cooling earn their keep for you. HP's Z2 Mini G1a is the enterprise version of the same idea if you need the support contract.
How it stacks up
Three comparisons, because Strix Halo isn't the only way to hold a big model at home.
- Versus a graphics card. A 5090 is far quicker on anything that fits in its 32GB, and it isn't close. But it can't hold a 70B model at all. Strix Halo can. It comes down to speed on small models against capacity for big ones.
- Versus the DGX Spark. NVIDIA's box has the same 128GB, a touch more bandwidth, and native CUDA, which a lot of AI tooling still wants. It's also Linux-only and dearer. Strix Halo is the x86 pick that runs Windows and costs less.
- Versus a Mac Studio. Apple's memory is much faster - two to three times the bandwidth - so a Mac answers quicker on the same model. It also costs a lot more, and it's macOS. If speed on big models is the point and you're happy on a Mac, the Studio wins. If value is, Strix Halo does.
The catches
- The NPU is mostly for show, for now. That 50-TOPS neural engine sounds great, but at the time of writing no local-LLM tooling makes real use of it. The graphics core does the work. Buy it for the memory, not the NPU.
- You'll fiddle with the software. AMD's ROCm stack still trails NVIDIA's CUDA, and owners often find the Vulkan backend faster than ROCm for llama.cpp. Ollama can even default to CPU-only until you set it straight. It works well once configured, but it's not plug-and-play.
- Check the memory mode. Some boxes ship with the memory in a mode that throttles bandwidth until you free it up in the BIOS - Beelink owners found exactly this. A one-time bit of fiddling, but do it, or you're leaving half the speed on the table.
None of that is a dealbreaker. It's the price of being early on a new class of machine. If you want 128GB for models in a small, quiet, x86 box, and you'll take a gentle pace to get it, Strix Halo is the best value going. Run your models through the memory calculator first, and if you decide you'd rather have speed than size, the GPU comparison is next door.