The cheapest machine for a 70B or gpt-oss-120b model, and which 128GB mini PC to buy

For a Prime Big Deal Days comparison, use the exact whole-PC cost worksheet to enter tax, shipping, required extras and any new membership cost. It covers the named shortlist configurations; recheck the exact seller, availability, returns and fit before counting a saving. The older prices below keep their original observation dates.

Based on maker and retailer listings, published specifications and third-party benchmarks, not our own testing.

Oct. 6 comparison: The current Prime local AI buying check has exact 128GB/2TB direct asks, GMKtec’s conflicting stock signals, Beelink’s approximately 35-day preorder and Framework’s out-of-stock 128GB option. The prices and rankings below are historical; compare the required model, context, software and complete delivered cost before treating a route as obtainable.

The short answer

For gpt-oss-120b, the lowest-priced complete 128GB machine in the Sept. 16 snapshot was a Ryzen AI Max+ 395 box: the GMKtec EVO-X2 with 128GB and a 1TB SSD, $3,499.99 at GMKtec’s store on Sept. 16. For a dense 70B, two used RTX 3090s cost less, $2,064 at the Sept. 12 eBay sold price, if the PC you already own has the power supply, slots and case room for them. Without that PC, the 64GB M5 Pro Mac mini at $2,699 holds a 70B with about 9GB to spare, on estimated speeds, with Apple’s announced Sept. 22 availability date; verify the current order’s delivery. The EVO-X2 was the lowest-priced complete route in that snapshot. Current orderability is unresolved, and benchmarks from a Framework Desktop or HP do not establish speeds on the GMKtec box.

The right size depends on the model you actually run, and the computer you own may already run a smaller one.

Routes for the EVO-X2 128GB: GMKtec’s store, which lists the cheaper 1TB version, or GMKtec EVO-X2 128GB/2TB on Amazon, check current price.

Route	Price, store and date	Holds a 70B, 39.1GB	Holds gpt-oss-120b, 65.4GB	Measured speed	Estimated speed	Next step
Keep what you own; run a smaller model	$0	Check yours	Check yours	Depends on your machine	—	What your computer runs
One 32GB card, in a PC you own	$1,299.99 Arc Pro B70 or $1,699.99 Radeon AI Pro R9700, Newegg, Sept. 16	No	Not outright	—	—	Which GPU runs your model
Two used RTX 3090s, in a PC you own	$2,064, two at GetPCParts’ $1,032 eBay sold price, Sept. 12	Yes, 48GB	Not outright; runs with expert layers in system RAM	70B: 16.29 t/s (XiongjieDai, May 2024); gpt-oss-120b, experts in RAM: 10-17 t/s (LLM Garage, Jan. 2026)	70B: 14-16 t/s (Compute Market)	Two cards or this box; two-card caveats: at least a 1,000W supply, slot spacing, case room
GMKtec EVO-X2, 64GB	$2,199.99 with 1TB, GMKtec, Sept. 16	Depends on the GPU share; not documented	No: 64GB installed	—	—	GMKtec’s store
M5 Pro Mac mini, 64GB	$2,699, Apple, Aug. 25; announced availability Sept. 22	Yes, about 9GB spare	No	None published	70B: 5-7 t/s	What each M5 Pro tier holds
GMKtec EVO-X2, 128GB	$3,499.99 with 1TB, $3,649.99 with 2TB, GMKtec, Sept. 16	Yes	Yes	70B: 4.5 t/s; gpt-oss-120b: 33.7 t/s; both on a Framework Desktop, a different machine with the same chip	—	GMKtec’s store, cheaper 1TB; EVO-X2 128GB/2TB on Amazon, check current price
Mac Studio M5 Max, 128GB	About $5,100, Daring Fireball, Aug., rounded	Yes	Yes	None published	70B: 11-14 t/s; gpt-oss-120b not estimated	Buying above $4,000

Measured speeds are single third-party runs, described in the speed section and on the linked pages. Estimates are either a named third party’s projection or bandwidth arithmetic, and a published measurement replaces them.

Two boxes would change the verdict if their prices or stock moved. The Framework Desktop’s $3,449 base price is $50.99 below the 1TB EVO-X2, but it is before storage, which Framework leaves configurable, and Framework listed it out of stock Sept. 16. Minisforum’s MS-S1 Max, $3,799 with 2TB that day, costs $149.01 more than the 2TB EVO-X2 and adds dual 10GbE, USB4 V2 and a PCIe x16 slot.

How much of the 128GB a model actually gets

The 128GB in these machines is system memory shared between the processor and the integrated Radeon 8060S. It is not VRAM on a card, and the share the GPU can address is a setting, not a spec.

On Windows the split is set through AMD’s Variable Graphics Memory, which reserves a block of system memory for graphics. HP documents the ceiling on its own product page for the Z2 Mini G1a: the unified memory architecture “allows you to assign up to 96 GB of memory exclusively to the GPU.” Framework’s Ryzen AI Max deep dive, published March 12, 2025, says the same thing and hints at the other path: “On Windows, up to 96GB can be dedicated to the GPU, and we’ve seen even higher numbers on Linux.”

On Linux the amdgpu driver’s graphics translation table lets the GPU map system memory on demand, so the BIOS framebuffer can stay small and the GTT size carries the model. In lhl’s Strix Halo benchmark notes, the test systems - pre-production Framework Desktops with 128GB of LPDDR5X-8000 - were set to 120GB of GTT through kernel module options. That is a contributor’s working configuration, not a vendor guarantee.

The 96GB figure above is HP’s and Framework’s documentation rather than AMD’s. GMKtec, Beelink and Minisforum state no maximum allocation on their store pages; GMKtec’s EVO-X2 page, read Sept. 16, states none for either its 64GB or its 128GB version. The tables on this page say “not documented” rather than assuming the 96GB number applies.

What that buys you in practice: gpt-oss-120b has a 65.4GB minimum and Llama 3.3 70B has a 39.1GB minimum at roughly 4-bit, so both sit inside a 96GB Windows allocation with room for context. A model that needs more than 96GB needs the Linux path. And 96GB to the GPU leaves 32GB for the operating system, the browser and whatever else is open.

Compare the machines

One row per 128GB machine we could price from a maker or retailer page on Sept. 15 or 16. Every machine here uses the same Ryzen AI Max+ 395, the same Radeon 8060S and the same 256GB/s memory bus, so the chip is not the variable.

Machine	Memory and storage	Documented max to the GPU	Price, store and date	Stock as we saw it	Route
Keep what you own, or rent by the hour	What you have	—	$0	—	Reasons not to buy
Framework Desktop	128GB LPDDR5X, soldered; storage configurable	96GB on Windows, “even higher” on Linux, per Framework	$3,449 at Framework, Sept. 16	“(Out of stock)” Sept. 15 and Sept. 16	Maker store only
GMKtec EVO-X2	128GB LPDDR5X-8000, 1TB or 2TB SSD	Not documented on GMKtec’s listing	$3,499.99 with 1TB and $3,649.99 with 2TB at GMKtec’s store, Sept. 16; $3,649.99 with 2TB on Newegg, sold direct by GMKtec, Sept. 15	Both listed as available at GMKtec Sept. 16, under a “Price increase coming soon” banner	Exact Amazon product, ASIN B0F53MLYQ6, the 2TB version
Minisforum MS-S1 Max	128GB LPDDR5X-8000, 2TB SSD	Not documented on Minisforum’s listing	$3,799 at Minisforum’s store, down from a $4,749 regular price, Sept. 16	Listed as available Sept. 16; the 4TB version showed “Estimated shipping time: Early October”	Exact Amazon product, ASIN B0G31Z6WQ7
Beelink GTR9 Pro	128GB LPDDR5X-8000, 2TB SSD	Not documented on Beelink’s listing	$4,349 at Beelink’s store, marked down from $4,699, Sept. 15	“Orders Ship within 35 Days”	Exact Amazon product, ASIN B0FPQQYWQ1
HP Z2 Mini G1a	Up to 128GB unified memory; the listing below is 128GB with 2TB	96GB to the GPU, per HP’s product page	$7,406.25 from antonline via Newegg, Sept. 15	Listed Sept. 15	Exact Amazon product, ASIN B0FSS64R69, which is the 128GB/1TB build

The HP is the outlier for a reason worth knowing before you shop it: it is sold as a commercial workstation through resellers, and reseller prices for the same machine scatter. Other Z2 Mini G1a listings on Newegg on Sept. 15 ran $4,599 and $6,167.10, but their memory configurations are unconfirmed, so they are not quoted here as 128GB prices.

The Bosgame M5 and the Corsair AI Workstation 300 also belong to this class; no current maker price is available for either.

Stock in this class moves. Treat every row above as what one store showed on the date given.

What 256GB/s buys, and what it doesn't

Token generation reads the model’s weights once per token, so memory bandwidth sets a ceiling. Divide 256GB/s by the weights and you get the arithmetic limit: for Llama 3.3 70B’s 39.1GB at roughly 4-bit, 256 divided by 39.1 is about 6.5 tokens per second. That is arithmetic, not a measurement, and real runs land under it.

They do. In lhl’s published llama.cpp runs on a 128GB Framework Desktop, Shisa V2 70B - a dense Llama-3 70B - generated 4.5 t/s on the HIP rocWMMA backend. Compute Market’s comparison expects 4-bit 70B generation to run at single-digit to low-double-digit tokens per second on both Strix Halo and Nvidia’s DGX Spark, because generation is bandwidth-bound; it cites no measured 70B generation figure.

The dense arithmetic does not transfer to mixture-of-experts models, and that is the case for buying one of these boxes. gpt-oss-120b’s 65.4GB would imply about 3.9 t/s by the same division, but it activates only a fraction of its parameters per token: lhl measured 33.7 t/s on the Vulkan AMDVLK backend, and StorageReview’s Z2 Mini G1a review reported close to 40 tokens per second on the same model with full GPU offload. If the models you care about are MoE, this class looks much better than its bandwidth does.

Prompt processing is the weak side, and that is the number people discover after they buy. Compute Market’s June 16 comparison puts prompt processing at roughly 340 tokens per second on a Strix Halo box against roughly 1,723 on the DGX Spark, about a 5x gap. lhl’s pp512 figures on the Framework Desktop were 94.7 t/s on the dense 70B and 430.6 t/s on gpt-oss-120b, with the note that “there’s a lot of performance still on the table when it comes to pp especially.” For scale, llama.cpp’s own gpt-oss guide records a used RTX 3090 processing gpt-oss-20b prompts at 5,170.56 t/s at 2,048 tokens - a different model on a different test, so not a like-for-like ratio, but the order of magnitude is the point. If you paste long documents or feed a coding agent large contexts, you will wait.

All of these are other people’s numbers on other people’s systems and builds. No per-machine prompt-processing figures we found separate the five boxes above from each other; the chip and the memory bus are the same, so expect the differences to come from cooling, drivers and configuration rather than from the badge.

Don't buy one of these if

The model already fits what you own. Check the model’s minimum against your machine’s unified memory or VRAM before spending anything. That check costs nothing, and the memory comparison prices every alternative per gigabyte if it doesn’t fit.

You need prompt-processing speed or CUDA. A used RTX 3090 or a new RTX 5060 Ti 16GB processes prompts far faster and runs the CUDA-only tooling these boxes can’t. The GPU guide matches models to 16GB, 24GB and 32GB cards with measured speeds, dated used prices and power-supply requirements.

You want macOS, or you were already pricing a Mac Studio. Used Ultras cost more per gigabyte but move memory about three times faster; the memory comparison puts them side by side with this class. Above $4,000 the usual buy-used advice inverts, and which machine to buy for local AI above $4,000 prices the Sept. 15 used Mac Studio asks against Apple’s new machines, where the new one is at or below the used ask at 64GB, 96GB and 128GB, and a used 256GB M3 Ultra asks about $800 less.

You’ll use it a few hours a month. Renting GPU time by the hour beats a $3,000 machine that idles. RunPod and Vast.ai both rent by the hour, and the break-even arithmetic prices this class against a memory-matched rented card. At the $3,499.99 price for the 1TB EVO-X2 above, spread over 36 months, the box costs $97.22 a month. At the $1.69-an-hour community rate for a 96GB RTX Pro 6000 that page carries, buying costs less only past about 58 hours a month, just under two hours a day. That page’s row uses the $3,649.99 2TB price, which works out to about 60 hours.

Where to buy

Framework Desktop, $3,449 before storage. Framework sells it only through its own store, which was showing “(Out of stock)” on Sept. 15 and again on Sept. 16.

GMKtec EVO-X2, 128GB. GMKtec’s own store listed the 1TB version at $3,499.99 and the 2TB version at $3,649.99 on Sept. 16, both as available, under a banner warning that a price increase is coming. GMKtec EVO-X2 128GB/2TB on Amazon, check current price. Newegg’s listing, sold direct by GMKtec at $3,649.99 for 2TB on Sept. 15. GMKtec’s store has the cheaper 1TB version.

Minisforum MS-S1 Max, 128GB and 2TB. Minisforum MS-S1 Max 128GB/2TB on Amazon, check current price. Minisforum’s own store had it at $3,799 on Sept. 16.

Beelink GTR9 Pro, 128GB and 2TB. Beelink GTR9 Pro 128GB/2TB on Amazon, check current price. Beelink’s own store listed it at $4,349 with orders shipping within 35 days.

HP Z2 Mini G1a. HP Z2 Mini G1a 128GB/1TB, Ryzen AI Max Pro 395, on Amazon, check current price. This is the 1TB build, not the 2TB one priced in the table. HP’s product page and Newegg’s Z2 Mini G1a listings; compare several resellers, because their prices for this machine do not agree.

Anything else in the class. Amazon search: Ryzen AI Max+ 395 128GB mini PCs. These machines ship in 32GB, 64GB, 96GB and 128GB configurations, so confirm 128GB on the exact listing before ordering.

Secondhand. eBay listings are worth a look given how far new prices have moved this year. A soldered-memory machine cannot be upgraded later, so confirm the configuration in the listing photos.

Still deciding between this class and everything else? Compare local AI memory by price per gigabyte and bandwidth, or match a model to a graphics card.