The best mini PCs for local AI inference in 2026

Start with the chip: five of the local-AI mini PCs in this guide run the same one. AMD's Ryzen AI Max+ 395 ("Strix Halo") with 128GB of unified LPDDR5X memory is the platform inside the Beelink GTR9 Pro, the GMKtec EVO-X2 and EVO-X3, the Framework Desktop, and HP's Z2 Mini G1a: same silicon, and therefore the same real-world model ceiling. The bandwidth figure you'll see quoted for these boxes, around 256GB/s, comes straight from AMD's published specs for the chip: a 256-bit memory bus running LPDDR5X-8000, which works out to 256GB/s. Every machine on this chip shares it. What separates those boxes is price, networking, cooling, repairability, and support, not how large a model they can hold. So most of this guide is about matching one of them to how you actually work.

The one thing that has changed: AMD's newer Ryzen AI Max+ PRO 495 raises the ceiling on the same design to 192GB, and machines built on it are now listed by their makers. They get their own section below.

Two machines break the pattern and belong on any 2026 shortlist. NVIDIA's DGX Spark (and the partner GB10 machines built on the same chip) brings the full CUDA stack to this class, and Apple's Mac Studio has more memory bandwidth than the AMD boxes: Apple specifies 614GB/s on the 40-core-GPU M5 Max and 1.2TB/s on the M5 Ultra, against roughly 256GB/s here, and bandwidth is what sets the speed on dense models. And if you searched for the best NUC for AI rather than a mini PC, there is a dedicated NUC section with two picks: an Intel NUC for integrated-graphics inference and a ROG NUC for CUDA.

One caveat colors every price below: 2026 is a bad year to buy memory. Micron told the SEC in mid-2026 that AI-driven demand is "outpacing industry supply", and that this has forced allocation decisions which "may impact certain customers and end markets". Treat the numbers as approximate and expect them to move.

Short version: for most people the GMKtec EVO-X2 is the best all-round 128GB Strix Halo box. It took the top spot from the Beelink GTR9 Pro on buyer data: when we checked Amazon in August 2026, the GTR9 Pro had 3.0 from 17 ratings, a third of them one-star, and the recurring complaints named its networking, the exact feature it was recommended for. Pick something else only if you need something specific: CUDA (DGX Spark), ECC and first-party support (HP Z2 Mini G1a), repairability (Framework Desktop), more than 128GB (a Ryzen AI Max+ PRO 495 machine), or Apple's bandwidth and MLX tooling (Mac Studio). For a NUC, the honest summary is that it is the right buy for models that fit in 64GB, or in a laptop GPU's 16GB, not for 70B-class work.

Quick picks (checked 23 October 2026):

---

GMKtec EVO-X2: Best overall#

Check price on Amazon↗

The EVO-X2 is our default recommendation into the 128GB Strix Halo tier, on two grounds: its price, and its buyer record. When we checked in August 2026 it had 4.2 out of 5 from 85 Amazon ratings, with buyers specifically reporting that it handles large local models. Same Ryzen AI Max+ 395, same unified-memory advantage as the other Strix Halo boxes.

Price check, 23 October 2026: the EVO-X2 is back on GMKtec's own store, which had stopped listing it in August. GMKtec lists it in 64GB and 128GB builds: $2,199.99 for 64GB with a 1TB SSD, and $3,499.99 (1TB) or $3,649.99 (2TB) for 128GB, on GMKtec's prices that day.

On price: on 23 October the 128GB EVO-X2 had the lowest price of any complete 128GB machine we could price-check, and it is still an expensive machine: this class costs multiples of a 32GB mini PC. Prices in this class change, so treat any figure you remember as possibly out of date. The buy link shows the current price for the configuration you pick.

Beyond the low price, its appeal is simply that it's a straightforward 128GB box: 2.5GbE networking and dual USB4 for fast external storage and displays. What it does not have is an OCuLink port for external-GPU expansion; that's a feature of GMKtec's pricier EVO-X3 (below), not this machine. Treat the EVO-X2 as the pure value play into 128GB, not as an eGPU platform.

The EVO-X3, if you want OCuLink: GMKtec's EVO-X3 runs the same Ryzen AI Max+ 395 with 128GB of LPDDR5X-8000 and adds an OCuLink (PCIe Gen4 x4) port for an external GPU. Despite the higher model number it has no higher model ceiling, and at GMKtec's own prices on 23 October it cost a little more than the EVO-X2: GMKtec listed it at $3,799.99 against $3,649.99 with a 2TB SSD, about 4% more. So the choice between them comes down to whether you want that OCuLink port.

Who it's for: Almost anyone who wants 128GB of Strix Halo: this is the default buy unless you need something on the list below.

Real trade-offs:

• The EVO-X2 had 4.2 out of 5 from 85 Amazon ratings when we checked in August 2026, but a rating count is not a support record: check GMKtec's warranty and return terms before you buy
• 2.5GbE networking, not the GTR9 Pro's dual 10GbE
• Same dense-model bandwidth ceiling as the rest of the Strix Halo field

When to pick something else: If you genuinely need dual 10GbE for always-on work, the Beelink GTR9 Pro builds it in, where the EVO-X2 and EVO-X3 list 2.5G LAN: read its reliability note first. If you want ECC memory and enterprise support, the HP Z2 Mini G1a.

---

Beelink GTR9 Pro: Dual 10GbE, with a reliability caveat#

Check price on Amazon↗

Read the buyer record before you buy this one. On paper the GTR9 Pro looked like the pick of the 128GB Strix Halo boxes, and it held our top spot on that basis: the same Ryzen AI Max+ 395 as the EVO-X2, plus dual 10GbE networking, dual USB4 and a vapor-chamber cooler with dual turbine fans. But when we checked its Amazon rating in August 2026 it was split rather than merely mediocre: 3.0 out of 5 from 17 ratings, with roughly a third at five stars and a third at one star, and the recurring complaints were about USB4/Thunderbolt and Ethernet connectivity. Those are precisely the features that justify paying more for this machine instead of the EVO-X2. Seventeen ratings is a small sample and it may well improve, but a split that shape can point to a fault that hits some units and not others, which is a different kind of risk from a machine that is simply unexceptional.

The silicon underneath is not in question. Its unified memory architecture is the point: CPU, GPU, and NPU all draw from the same pool of up to 128GB LPDDR5X, so a large model sits entirely in fast memory instead of being shuttled between VRAM and system RAM. There is a ceiling inside the ceiling worth knowing: HP's data sheet for the Z2 Mini G1a says up to 96GB of the 128GB can be assigned to the GPU, and GMKtec's own listing for this silicon says the same, "Up to 96GB VRAM".

Be realistic about speed. Memory bandwidth (~256GB/s) is the ceiling, and it bites hardest on dense models. A dense 70B model at Q4 runs at low single-digit tokens per second here: usable for batch work, slow for interactive chat. Where the chip shines is smaller and sparse models: small dense models (7B–13B) run at tens of tokens per second, and 30B-class mixture-of-experts (MoE) models are much faster still, because only a fraction of their weights are active per token. Even a large MoE like gpt-oss 120B stays usable rather than crawling. Understand the dense-vs-MoE distinction before you buy: it changes how fast the same box feels.

On price: the GTR9 Pro sits above the EVO-X2 for the same silicon and the same 128GB: you are paying for the networking and the cooling, not for more capability. On 23 October Beelink's own store listed it at $4,349, in one build (128GB with a 2TB SSD).

Who it's for: Someone who specifically needs dual 10GbE and a vapor-chamber cooler, has read the reliability note above, and is buying from a storefront with a return policy they are willing to use.

Real trade-offs:

• A split reliability record, with complaints aimed at the networking and USB4 that are its reason to exist
• Same ~256GB/s bandwidth and model ceiling as every other Strix Halo box: you're paying for I/O and cooling, not more capability
• Overkill if your daily models are 7B–13B: you'd be buying headroom you won't use

When to pick something else: For most people, the GMKtec EVO-X2: same chip, same 128GB, and 4.2 out of 5 from 85 Amazon ratings when we checked in August 2026, against the GTR9 Pro's 3.0 from 17. If you want repairability and documentation, the Framework Desktop.

---

Framework Desktop: Best supported / repairable#

Framework's pitch is hardware you repair rather than throw away. It's the same Ryzen AI Max+ 395 with up to 128GB LPDDR5X, but with Framework's repairability ethos and documentation. Its 128GB build has the same capacity as the rest of this tier, so it holds the same 70B-class and large-MoE models, with the same bandwidth-bound caveats on dense models.

Framework sells it direct as a DIY Edition in three builds: 32GB on the Ryzen AI Max 385, and 64GB or 128GB on the Ryzen AI Max+ 395. On Framework's configurator on 23 October the 128GB build listed at $3,449 before storage and an OS, about the same as GMKtec's 128GB EVO-X2, so buy this one for the support-and-parts story rather than to save money. Framework also lists a 192GB version on the newer Ryzen AI Max+ PRO 495, covered in the 192GB section below.

Who it's for: Buyers who want the 128GB Strix Halo platform from a vendor whose pitch is repairability and documentation.

Real trade-offs:

• About the same price as the 128GB EVO-X2 before you add storage and an OS, for the same core silicon and model ceiling: compare configured totals, not the board price
• The memory is soldered LPDDR5X, so "repairable" applies to storage, fans, and I/O, not the RAM

When to pick something else: If you want a finished box with the SSD included, the EVO-X2. If you specifically need ECC memory and first-party workstation support, the HP Z2 Mini G1a.

---

HP Z2 Mini G1a: Best first-party workstation#

Check price on Amazon↗

This is the real answer to "I want a name-brand workstation, not a white-box mini PC." The Z2 Mini G1a runs the Ryzen AI Max+ PRO 395 with up to 128GB of LPDDR5x ECC memory and Radeon 8060S graphics, and its 128GB is enough memory to hold gpt-oss 120B without a discrete GPU. HP sells it configured to order, and we could not price-check a 128GB build on 23 October, so compare a configured quote against the consumer boxes: what you are paying for here is ECC and a first-party support contract.

An earlier version of this guide named an "AMD Ryzen AI Halo Developer Platform" here, and a later one removed it because we could not find it as a product. It exists now: AMD lists a Ryzen AI Halo Developer Platform in a 128GB Ryzen AI Max+ 395 version and a 192GB Ryzen AI Max+ PRO 495 version, sold through Micro Center. AMD positions it as a developer box rather than a workstation, so it sits in the 192GB section below, and the Z2 Mini G1a stays the pick when ECC memory and an HP support contract are the point.

Who it's for: Teams and professionals who need ECC memory, first-party warranty and support, and the accountability of an HP workstation SKU.

Real trade-offs:

• The same Ryzen AI Max+ core and the same ~256GB/s ceiling as the consumer boxes; you're paying for ECC, support, and brand
• The PRO chip and ECC matter for reliability, not for raw inference speed
• If provenance and support aren't requirements, the consumer Strix Halo boxes do the same work

When to pick something else: If you don't need ECC or enterprise support, any of the consumer Strix Halo boxes. If you need CUDA, the DGX Spark.

---

Ryzen AI Max+ PRO 495 machines: 192GB, for models that outgrow 128GB#

The memory ceiling has moved. AMD's Ryzen AI Max+ PRO 495 ("Gorgon Halo") keeps the 395's 16 Zen 5 cores and 40-CU graphics (now called Radeon 8065S) but supports up to 192GB of LPDDR5X-8533 on the same 256-bit bus, and AMD lists its own 192GB system at 273GB/s: a small step up from the 395 boxes' ~256GB/s, and the same figure NVIDIA gives the DGX Spark. What you are buying is capacity. AMD rates its 192GB box for models up to 300B parameters, against 200B for its 128GB version, and Minisforum and GMKtec both say up to 160GB of their 192GB machines can be used as VRAM.

We found five machines on it:

• ACEMAGIC F9A (Max+ PRO 495, 192GB): a 2TB SSD, dual 2.5GbE, USB4 and OCuLink, listed at $6,499 on ACEMAGIC's store on 23 October.
• AMD Ryzen AI Halo Developer Platform (Max+ PRO 495, 192GB): AMD's own box, with 10GbE, a 2TB SSD and a Linux developer image, sold through Micro Center.
• Framework Desktop (Max+ PRO 495, 192GB): listed at $6,799 on Framework's configurator on 23 October.
• GMKtec EVO-X5 Pro (Max+ PRO 495, 192GB): dual 10GbE and two USB4 v2 ports, listed at $6,799 with a 2TB SSD or $7,099 with 4TB on GMKtec's store on 23 October.
• Minisforum MS-S1 MAX-P495: 192GB with a 2TB SSD, dual 10GbE and USB4 v2, listed at $7,399 on Minisforum's store on 23 October.

Who it's for: People whose target model genuinely does not fit in 128GB. Bandwidth barely moves, so a model that already fits on a 128GB Strix Halo box will not run meaningfully faster here.

Real trade-offs:

• Framework's and GMKtec's 192GB builds list at close to twice the price of their 128GB builds: you pay nearly double for half again the memory
• The same RDNA 3.5 graphics and nearly the same bandwidth: more room, not more speed
• A new platform with a short track record; buy from a seller whose return policy you would actually use

When to pick something else: If your models fit in 128GB, an EVO-X2 or another 395 box does the same work, and the 128GB EVO-X2 lists at about half the price of GMKtec's 192GB EVO-X5 Pro. If you need CUDA, the DGX Spark.

---

NVIDIA DGX Spark: Best for CUDA#

Check price on Amazon↗

If NVIDIA's software stack is non-negotiable, the DGX Spark is the machine. It pairs the GB10 Grace Blackwell superchip with 128GB of coherent unified LPDDR5x at 273GB/s, NVIDIA's stated up to 1 petaFLOP of FP4 compute, and the full CUDA stack. NVIDIA rates a single 128GB Spark for models up to 200B parameters, and up to four Sparks can be linked over its 200Gbps ConnectX-7 port for larger ones.

We could not read NVIDIA's current price for the 128GB Spark on 23 October, so check the buy link above. NVIDIA's page lists GB10 systems from Acer, ASUS, Dell, Gigabyte, HP, Lenovo and MSI under their own names. It also lists a 64GB version, rated for models up to 100B parameters and sold only through partners: NVIDIA said on 2 October that it would go on sale from Acer, ASUS, Dell, Gigabyte, HP and MSI on 23 October, starting at $4,999. That is half the memory of a 128GB EVO-X2 at a higher starting price than GMKtec's $3,499.99, so buy the 64GB Spark for CUDA, not for capacity.

A related launch, RTX Spark: NVIDIA said on 17 October that laptops built on its RTX Spark chip, which pairs a Grace CPU with Blackwell RTX graphics and up to 128GB of unified memory, were launched by ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,399. Those are laptops. ASUS has also announced a ProArt GR1X mini PC on the same chip (the RTX Spark N1X), with up to 128GB of LPDDR5X, a 10G LAN port and Windows, but on 23 October its page offered only a sign-up and no price, so it is not a pick here yet.

Who it's for: Developers and researchers whose workflows are built on CUDA, and anyone who wants to run models up to ~200B parameters on a first-party NVIDIA box without a data-center GPU.

Real trade-offs:

• The 128GB Spark was not price-checked on 23 October: compare it on the day against a 128GB Strix Halo box
• CUDA and NVIDIA's ~1 PFLOP FP4 figure are the draw; like every unified-memory box in this class, dense-model token rates are still gated by memory bandwidth
• Overkill unless you specifically need CUDA

When to pick something else: If you don't need CUDA, a 128GB Strix Halo box does the same local-LLM job; compare prices on the day. If you want more memory bandwidth, look at Apple.

---

Apple Mac Studio (and Mac Mini): Best Apple option#

Check price on Amazon↗

The list would be incomplete without Apple, and the bandwidth is why. The Mac Studio with the 40-core-GPU M5 Max runs unified memory at 614GB/s (the base 32-core-GPU M5 Max is 460GB/s), and Apple lists the M5 Ultra at 1.2TB/s: more than double and nearly five times a Strix Halo box's ~256GB/s. Apple refreshed the line on August 25, 2026. As of 23 October, Apple's specs page lists the M5 Max from 36GB, with the 48GB, 64GB and 128GB options tied to the 40-core-GPU chip, and the M5 Ultra from 96GB, with the 256GB and 512GB options tied to the top 36-core-CPU, 80-core-GPU chip. That pairing matters before you configure: the big memory options force the top chip in each line. The Mac Mini is a legitimate entry point for smaller models. On the software side, Apple's MLX framework and LM Studio both run local LLMs on Apple silicon.

On dense models memory bandwidth is the bottleneck, so the M5 Max's 614GB/s, about 2.4 times a Strix Halo box's ~256GB/s, is where its speed advantage comes from, and the M5 Ultra's 512GB option opens up models the 128GB boxes can't hold at all.

On 23 October Apple's U.S. store listed the Mac Studio from $2,499 with the M5 Max and from $5,499 with the M5 Ultra, the same starting prices Apple announced in August. The bigger memory options need the top chip in each line, so price the exact build you want on Apple's store. Treat Apple as the bandwidth-and-capacity play.

Who it's for: People already in macOS, anyone who wants more memory bandwidth than the AMD boxes for dense-model inference, and researchers who need the largest unified-memory pool Apple offers (M5 Ultra).

Real trade-offs:

• No CUDA: some frameworks and models are CUDA-first and need porting or run through MLX/Metal
• The 256GB and 512GB options need the top M5 Ultra, and we did not price-check those builds on 23 October
• macOS-only; if your stack assumes Linux/Windows, factor the switch

When to pick something else: If your tooling is CUDA-bound, the DGX Spark. If you want 128GB on Windows or Linux, a Strix Halo box.

---

Minisforum AI X1 Pro-470: Best budget starter#

The AI X1 Pro-470 is built on AMD's Ryzen AI 9 HX 470 (Gorgon Point, a Zen 5 refresh of Strix Point) with Radeon 890M graphics. On 23 October Minisforum's store listed it at $727 as a barebone, $1,327 with 32GB and a 1TB SSD, $1,655 with 64GB and 1TB, and $2,039 with 96GB and 2TB. Its memory is not soldered: it takes two DDR5-5600 SO-DIMMs up to 128GB, and Minisforum sells 32GB, 64GB and 96GB builds.

That upgrade path needs one honest qualifier. Two SO-DIMMs are a much narrower path to memory than Strix Halo's 256-bit bus, so a 96GB AI X1 Pro can hold models a 32GB box cannot, but it will run them noticeably slower than a 128GB Strix Halo machine. Buy the 32GB or 64GB build for 7B–30B work, and treat extra capacity as headroom rather than as a cheap route to 70B.

One bonus at this price: it includes an OCuLink port, so you can attach an external NVIDIA GPU for CUDA workloads later, the same eGPU escape hatch GMKtec adds on the EVO-X3.

Who it's for: Developers getting started with local AI inference, and power users whose daily models are sub-30B, who want AMD's Gorgon Point architecture without paying for a 128GB Strix Halo box.

Real trade-offs:

• Its DDR5 SO-DIMMs are upgradeable but slow next to Strix Halo's soldered LPDDR5X: more capacity does not buy more speed
• Its integrated GPU has meaningfully less compute than the Max+ 395; large-model inference speed reflects that
• Thermals under sustained AI load are worth checking in reviews before you buy

When to pick something else: If you regularly need 70B-class models at usable speeds, save for a 128GB Strix Halo box. The gap here is both capacity and bandwidth.

---

Geekom A9 Max: Best complete 32GB package#

Check price on Amazon↗

On 23 October Geekom's own store listed the A9 Max in three builds: $1,299 for the Ryzen AI 9 HX 370 with 32GB and a 1TB SSD, $1,499 for the same chip with 2TB, and $1,599 for a newer build on the Ryzen AI 9 HX 470 (the same chip as the Minisforum above) with 32GB and 2TB. That $1,299 build was the lowest list price of any complete machine we could price-check for this guide, just under Minisforum's 32GB AI X1 Pro-470 but on the older chip, and Geekom says it comes with Windows 11 Pro installed.

The HX 370 is still a capable chip for local AI. It handles 7B–13B models well, and the A9 Max ships as a complete box with the SSD included.

Who it's for: Buyers who want a finished 32GB machine from one vendor, with the SSD and Windows included. If the HX 470 build is the one you want, compare it directly with the AI X1 Pro-470 on the same chip.

Real trade-offs:

• On the same HX 470 chip, Geekom's build listed at $1,599 with 2TB on 23 October, against $1,327 for Minisforum's 32GB AI X1 Pro-470 with 1TB
• The HX 370 builds use an earlier Strix Point chip; check which processor a listing actually carries
• Plan to bring your own local-AI software stack

When to pick something else: For the HX 470 chip at a lower list price, plus an OCuLink port for a later external GPU, the Minisforum AI X1 Pro-470. For 70B-class models, a 128GB Strix Halo box.

---

ASUS NUC 16 Pro and ROG NUC: the best NUCs for local AI, and when a NUC is the wrong answer#

A NUC, the small-PC line ASUS now makes, answers a different question from the Strix Halo boxes above. The 128GB AMD machines exist to hold 70B-class models with room to spare, and no NUC here is built for that. What a NUC does well is run models that fit in 64GB of system memory or in a laptop GPU's VRAM, on Intel hardware, with ASUS's support behind it. Three are worth knowing, and they suit different buyers.

ASUS NUC 16 Pro (Core Ultra X7 or X9): Best NUC for local AI#

The NUC 16 Pro is ASUS's NUC on Intel's Core Ultra Series 3 (Panther Lake). The version to buy for AI is the one with a Core Ultra X7 358H or X9 378H. Those carry Intel's Arc B390 graphics with 12 Xe-cores, which Intel rates at 122 peak INT8 TOPS for the GPU and 50 for the NPU on the X7 358H, and they come with soldered LPDDR5X: ASUS's spec sheet lists LPDDR5X-8533 in 32GB and 64GB builds. Get the 64GB build: it is double the NUC 14 Pro AI's ceiling, and memory capacity, not TOPS, decides which local LLMs fit.

Read the listing carefully, because the NUC 16 Pro comes in two very different shapes. The Core Ultra 7 356H and Core Ultra 5 versions take DDR5 CSO-DIMMs up to 128GB, which looks like the better deal on paper. Intel's own spec sheet shows why it is not, for AI: the 356H has 4 Xe-cores against the X7's 12. More memory on a much weaker GPU will hold a larger model and run it slowly.

Who it's for: Anyone who wants a NUC specifically, runs models up to the 30B class, and values a compact Intel/Windows box with dual 2.5GbE, two Thunderbolt 4 ports and ASUS support over raw model-size headroom.

ASUS ROG NUC (2025) with RTX 5080 Laptop GPU: Best NUC for CUDA#

If what you need is NVIDIA's software stack in NUC form, this is the one. ASUS builds the ROG NUC (2025) with up to an Intel Core Ultra 9 275HX and a choice of GeForce RTX 5060, 5070, 5070 Ti or 5080 Laptop GPUs, and NVIDIA lists the RTX 5080 Laptop GPU with 16GB of GDDR7. That is the whole trade: models that fit inside 16GB of dedicated VRAM run at full CUDA speed, and anything larger spills into system RAM (up to 96GB of DDR5 here) and slows sharply. Buy the RTX 5080 build for AI; the 8GB and 12GB GPUs below it cut the model ceiling hard.

Who it's for: CUDA-first developers whose daily models fit in 16GB of VRAM, who want a compact NVIDIA box, and anyone who also games on the same machine.

ASUS NUC 14 Pro AI: the previous pick, now outclassed#

The NUC 14 Pro AI was this guide's Intel pick. It runs Intel's Core Ultra Series 2 (Lunar Lake) with LPDDR5X-8533 in 16GB or 32GB builds, and ASUS rates it at "up to 48 NPU TOPS" and "120 platform TOPS", a good illustration of why this guide keeps telling you to discount TOPS: none of it changes how big a model a 32GB ceiling will hold. ASUS still lists it, and it is still a capable small-model machine, but the NUC 16 Pro with 64GB now does what it does with twice the memory.

Real trade-offs, for every NUC here:

• No NUC here reaches 128GB of fast unified memory: 70B-class models belong on a Strix Halo box, a DGX Spark or a Mac Studio
• The ROG NUC's CUDA speed stops at 16GB of VRAM; past that, it is a slow machine with a fast GPU

When to pick something else: If running 70B-class models is a current or near-term goal, the AMD Strix Halo options are the better platform. If you want CUDA and 128GB in one box, the DGX Spark.

---

Comparison#

Why this table ranks cost instead of listing prices. Prices in this class change, and a printed figure can go out of date before the article is updated. This table records relative position on 23 October, which is what you need in order to choose. For the number itself, use the buy link, which shows the live price.

---

How to choose#

1. Capacity sets the ceiling; bandwidth sets the speed. Marketing loves TOPS. What determines which models you can run is unified memory capacity: 32GB handles 7B–13B (and 30B quantized); 128GB is the target for 70B-class work, and 192GB is now an option for models beyond that. But capacity isn't speed: memory bandwidth is. That's why Apple's 40-core-GPU M5 Max (614GB/s) and the M5 Ultra (1.2TB/s) have a higher speed ceiling on dense models than any Strix Halo box (~256GB/s) at the same capacity. Decide capacity first, then bandwidth.

2. Understand MoE vs. dense: it decides how fast a model feels. The same 128GB Strix Halo machine runs a dense 70B model at low single-digit tokens per second, but a 30B-class MoE runs much faster, and even a large MoE like gpt-oss 120B stays usable. Sparse (MoE) models activate only a fraction of their weights per token, so they run much faster on these bandwidth-limited boxes than dense models of similar size. If your target models are MoE, these machines are much faster than the dense-70B numbers suggest, and the reverse is just as true.

3. The Ryzen AI Max+ 395 is the x86 value pick, not the outright best. On 23 October the lowest-priced complete 128GB machine we could price-check ran this chip. But "best overall" depends on the axis: Apple's 40-core-GPU M5 Max has more than double the memory bandwidth and the M5 Ultra nearly five times, and NVIDIA's GB10 (DGX Spark) adds CUDA and, by NVIDIA's figure, up to 1 petaFLOP of FP4 compute. Strix Halo's edge is price on the machines we could check, not every metric.

4. Factor the 2026 memory shortage. Micron's fiscal third-quarter 2026 10-Q describes AI-driven growth outpacing industry supply, and says robust demand and constrained supply "has led to increased pricing" across its portfolio. Every price here is approximate and moving: check current listings, and don't assume last quarter's number holds.

5. Know your model ceiling before you buy. If your real workflow is an 8B coding assistant, you don't need a 128GB Strix Halo box: a 32GB Minisforum or Geekom is fine, for less than half the price of a 128GB EVO-X2 on the makers' list prices on 23 October. If you want to run 70B-class or large-MoE models, you need a 128GB box. And if you need CUDA or ECC, that decides the machine before price does. Be honest about which tier you actually need.

6. Consider the eGPU escape hatch. Boxes with an OCuLink port (the budget Minisforum AI X1 Pro-470, or GMKtec's pricier EVO-X3) let you attach an external NVIDIA GPU for CUDA acceleration later. If you might need CUDA but don't want to buy a DGX Spark today, that expansion path is a real hedge. (The GMKtec EVO-X2 does not have OCuLink; the EVO-X3 does.)

---

The verdict#

For most people in 2026, the GMKtec EVO-X2 is the safe default: the lowest-priced complete 128GB Strix Halo box we could price-check on 23 October, with 4.2 out of 5 from 85 Amazon ratings when we checked in August 2026, and the EVO-X3 is GMKtec's pick on the same chip if you want an OCuLink port. Want dual 10GbE and a vapor-chamber cooler for always-on work? The Beelink GTR9 Pro offers both on 128GB: read its reliability note above first. Want repairability and support? The Framework Desktop. Need more than 128GB? One of the new 192GB Ryzen AI Max+ PRO 495 machines. Came for a NUC? The ASUS NUC 16 Pro with a Core Ultra X7 or X9 and 64GB, or the ROG NUC with an RTX 5080 Laptop GPU if you need CUDA. Pick something else only for a specific reason: CUDA at 128GB (DGX Spark), ECC and first-party support (HP Z2 Mini G1a), or Apple's bandwidth and huge memory pool (Mac Studio). And if your real daily model stays under 13B, none of that applies: a 32GB Minisforum AI X1 Pro-470 does the job for well under half the price of a 128GB EVO-X2 on the makers' list prices on 23 October, and a 32GB Geekom A9 Max also costs under half as much. Whatever you pick, buy for the model type you actually run, MoE or dense, because that decides how fast it feels as much as any spec sheet does.

---

Our picks#

🏆 Top pick: GMKtec EVO-X2 (best overall). Ryzen AI Max+ 395 with up to 128GB unified LPDDR5X (~256GB/s): the same silicon as the other Strix Halo boxes at the lowest price of any complete 128GB machine we could price-check, and 4.2 out of 5 from 85 Amazon ratings when we checked in August 2026.