Beelink GTR9 Pro vs GMKtec EVO-X2 in 2026: Same Chip, Same 31 tok/s — So You're Buying Cooling, 10GbE, and a Price Roll

By RunAIHome Team · 12 min read
mini-pcstrix-haloryzen-ai-maxbeelinkgmkteclocal-llmhardwarecomparison

TL;DR: Both boxes run the same Ryzen AI Max+ 395 with 128GB of LPDDR5X-8000 at 256 GB/s, so LLM speed is a tie — roughly 31 tok/s on gpt-oss 120B on either machine. The GTR9 Pro buys you dual 10GbE and a vapor-chamber cooler that holds AI loads at 39–41 dBA; the EVO-X2 buys you lower idle power and, usually, a lower price. The GTR9 Pro’s early dual-10GbE boards shipped with a genuine Ethernet-under-load defect, since fixed by firmware and a board revision.

Beelink GTR9 ProGMKtec EVO-X2Used RTX 3090 build
Best for128GB inference + fast home network128GB inference at the lowest price of the weekEverything that fits in 24GB
Price (Sep 2026)~$1,899–$1,999 street (Amazon); much higher direct~$2,199 direct / $2,299 Micro Center (Jun 2026, drifting)~$1,070 card, ~$1,400 build
gpt-oss 120B31.41 tok/s~31 tok/swon’t fit
Networkingdual 10GbE + Wi-Fi 7dual 2.5GbE + Wi-Fi 7whatever your board has
The catchv1.0 boards had a real 10GbE-under-load bug4× slower wired networkinghard 24GB ceiling

Honest take: These are the same computer where it counts, so buy whichever 128GB/2TB config is cheaper the week you order — that’s usually the Beelink GTR9 Pro on Amazon right now. Pay an EVO-X2 premium only if the GTR9 Pro’s networking history spooks you; pay a GTR9 Pro premium happily if you move models over your LAN, because dual 10GbE is the one spec the EVO-X2 can’t answer.

Why both boxes generate tokens at the same speed

The inference numbers are interchangeable because the silicon is interchangeable. Both machines use the AMD Ryzen AI Max+ 395 — “Strix Halo,” the same 16-core Zen 5 APU with a Radeon 8060S iGPU (40 RDNA 3.5 compute units) we covered in our Strix Halo deep dive — and both solder the same 128GB of LPDDR5X-8000 onto a 256-bit bus for 256 GB/s of bandwidth. Token generation on a machine like this is memory-bandwidth-bound, so two boxes with identical memory systems decode at identical speeds, whatever their cases look like.

The measurements bear that out. ServeTheHome ran gpt-oss 120B (MXFP4, LM Studio on Windows) on the GTR9 Pro and got 31.41 tok/s at 125–128W of wall draw. The EVO-X2 posts ~31 tok/s on the same model in independent testing. On dense models the story repeats: Llama 3.3 70B decodes at ~5 tok/s on both (we measured the same on the EVO-X2 in our full review), and community llama.cpp tuning on Linux pushes gpt-oss 120B into the mid-40s to low-50s tok/s on Strix Halo generally — again, chip behavior, not box behavior.

So a spec-sheet benchmark shootout is the wrong way to pick between these two. What you’re actually choosing is everything wrapped around the chip: the cooler, the network controller, the idle power, and — in 2026’s memory market — whichever price tag hasn’t spiked that week.

Cooling and noise: the vapor chamber earns its keep, barely

The GTR9 Pro’s headline hardware difference is a full-coverage vapor chamber with dual turbine fans that fills most of the chassis — Beelink claims it holds 140W TDP at 32 dB. Measured reality is close: ServeTheHome recorded 36–37 dBA at idle (in a studio with a 34 dBA noise floor) and 39–41 dBA running gpt-oss 120B at about 120–128W system draw. That’s a machine you can keep on the desk while a 120B-parameter model streams tokens next to you.

The EVO-X2 cools the same chip with three copper heat pipes, two CPU fans, and a third fan dedicated to the M.2 drives. Tom’s Hardware and ServeTheHome measured 37–38 dBA at idle and 41–43 dBA under Llama 3.3 70B Q6_K inference at 147–160W draw. Caveat before you crown a winner: the two load tests aren’t identical — a dense 70B run pulls ~25W more than a gpt-oss 120B MoE run, and more watts means more fan. Under comparable AI loads the honest summary is: the GTR9 Pro is 2–4 dBA quieter, which is noticeable side by side and irrelevant from across the room.

One asterisk on the Beelink: at least one owner has reported fan noise at 52 dBA — an outlier against multiple reviewer measurements, but a reminder that mini PC acoustics have unit-to-unit variance. If yours screams, that’s a support case, not the design.

Idle power flips the table. The EVO-X2 idles at 8–14W in most testing; the GTR9 Pro idles at 16–25W depending on what Windows 11 is doing. If the box runs 24/7 as a home AI server — the way most people deploy these — idle hours dominate the power bill, as we worked through in our 24/7 server power math. At the ~10W difference and $0.12/kWh, that’s about $10/year: real, not decisive.

Networking: the one spec gap — and the one real defect

This is the only place the two boxes genuinely differ in capability. The GTR9 Pro ships dual 10GbE plus Wi-Fi 7 and two USB4 ports; the EVO-X2 ships dual 2.5GbE (Realtek 8125BG) plus Wi-Fi 7 and two USB4 ports. If you pull models off a NAS, serve inference to other machines on your LAN, or shuttle 60GB+ GGUF files around, 10GbE moves them 4× faster — a 40GB model file lands in under a minute instead of ~2.5 minutes, every time you switch models.

But the GTR9 Pro’s 10GbE has a history you should know before buying. Early v1.0 boards used Intel’s E610 controller, and under sustained load the Ethernet would malfunction — link drops and BSODs documented across a 27-page thread on Beelink’s own forum. The working fix that emerged from Beelink support was a firmware-and-driver stack, not a workaround:

System BIOS:          v1.08
NIC firmware:         Intel E610 NVM 1.30
Ethernet driver:      Intel Ethernet Driver Pack 30.4

Owners who applied all three report the BSODs and drops stopped. Beelink also quietly revised the hardware: v2.2 boards replaced the Intel E610 with Realtek’s RTL8127 10GbE controller and improved power circuitry. The product page still just says “dual 10GbE” without naming the controller, so if you buy one, check which board you received:

lspci | grep -i ethernet
# v1.0 board  -> Intel Corporation ... E610 entries
# v2.2 board  -> Realtek ... RTL8127 entries

If lspci (or Windows Device Manager) shows the Intel E610, update the BIOS, NIC firmware, and driver to the versions above before you trust it with an overnight job on wired networking. If it shows the RTL8127, you have the revised board. This episode is the strongest argument the EVO-X2 has in this comparison: its 2.5GbE is slower, but it’s been boring — and boring is a feature in an always-on server.

Storage, ports, and the rest of the box

Storage is a wash. Both machines take two M.2 2280 PCIe 4.0 drives for up to 16TB total and both ship the 128GB flagship config with a 2TB SSD — the GTR9 Pro’s is a Crucial drive rated to 7,000MB/s. The practical difference is serviceability: the EVO-X2’s recessed M.2 bays are fiddlier to reach than the Beelink’s. Both have two USB4 (40Gbps) ports, HDMI 2.1, and Wi-Fi 7 with Bluetooth 5.4; the GTR9 Pro adds an SD card slot and DisplayPort 2.1, which matter to photographers and multi-monitor setups and to nobody else.

Neither machine lets you touch the memory — the LPDDR5X is soldered on both, which is exactly why the 128GB config is the only one worth buying. If 128GB sounds like more than your models need, check your target models against our VRAM calculator before spending this kind of money; a used RTX 3090 at ~$1,070 is roughly 3× faster on everything that fits in 24GB.

Price: the least stable spec on either machine

Every price in this section moves with the DRAM market, so treat these as dated data points rather than gospel. What’s verifiable as of September 2026:

  • Beelink GTR9 Pro (128GB/2TB): launched at a $1,985 pre-order price in 2025. Amazon US listings have run $1,899–$1,999 through 2026, with periodic dips below $1,900. Beelink’s own direct store has ballooned during the memory crunch — the Canadian storefront showed CA$4,349 in late August 2026 — so Amazon is currently the sane place to buy.
  • GMKtec EVO-X2 (128GB/2TB): launched near $1,999, spiked to roughly $3,200 in late 2025, and sat at about $2,199 direct / $2,299 at Micro Center as of June 2026. The $1,499 figure that still floats around forums was the 64GB/1TB review config at launch — no 128GB machine has sold for that.

The pattern to internalize: these two products swap places on price as memory contracts reprice, and the delta has been as wide as $400 in either direction. The chip inside doesn’t change. So the buying rule for this segment is genuinely check both prices the week you order and take the cheaper one — with the two tiebreakers below.

Who should buy which

Buy the GTR9 Pro if you’ll use the network. A home lab with a NAS, a 10GbE switch, or a second inference box gets real value from dual 10GbE — it’s the difference between the mini PC being a peer on a fast network and being an island. Verify the board revision on arrival (or apply the BIOS v1.08 + NVM 1.30 + driver 30.4 stack on an Intel E610 unit), and you get the quieter cooler as a bonus. At its recent $1,899–$1,999 Amazon pricing, it’s also been the cheaper box, which makes this an easy call.

Buy the EVO-X2 if you want the box that’s had no drama. Same speed, lower idle draw, a networking stack with no forum saga, and GMKtec has been shipping Strix Halo boxes since early 2025 — it’s the incumbent for a reason. If its price dips back under the GTR9 Pro’s, it’s the default again. Our full EVO-X2 review covers what living with it is like.

Buy neither if your models fit in 24GB. Both boxes exist for one job: holding models a consumer GPU can’t. If you mostly run 7B–32B models, a used RTX 3090’s 936 GB/s decodes ~3× faster for half the money. And if you’re not sure whether your workload justifies $2,000 — that’s the cheapest question in this article to answer: rent a big-memory GPU on RunPod for an evening and see which model sizes you actually reach for. We compared the premium tier of this segment — DGX Spark and AMD’s own dev kit — in our 128GB dev-kit head-to-head, and the broader budget landscape in our mini PC guide. If the box will serve a local coding stack, the tooling side is covered at aicoderscope.com.

What to actually buy

Prices as of September 2026, all taken from the comparison above:

Your situationThe machinePriceWhere
100B+ MoE models at home, plus a fast LAN (NAS, 10GbE switch)Beelink GTR9 Pro (128GB/2TB)~$1,899–$1,999Check price
Same workload, but you want the box with zero networking historyGMKtec EVO-X2 (128GB/2TB)~$2,199–$2,299Check price
Everything you run fits in 24GBUsed RTX 3090~$1,070Check price
Not sure 128GB is your workload yetRented GPU, ~$1/hrpay per hourRunPod

FAQ

Is the Beelink GTR9 Pro faster than the GMKtec EVO-X2 for local LLMs? No. Both use the Ryzen AI Max+ 395 with 128GB LPDDR5X-8000 at 256 GB/s, and both decode gpt-oss 120B at ~31 tok/s. Token generation is bandwidth-bound, and the bandwidth is identical.

Was the GTR9 Pro Ethernet problem fixed? Yes, two ways. On original v1.0 boards (Intel E610 controller), the combination of BIOS v1.08, NIC firmware NVM 1.30, and Intel driver pack 30.4 stopped the under-load drops and BSODs owners documented on Beelink’s forum. Revised v2.2 boards replaced the E610 with a Realtek RTL8127 and don’t exhibit the issue.

Which box is quieter? Measured under AI load, the GTR9 Pro: 39–41 dBA running gpt-oss 120B (~120–128W) versus the EVO-X2’s 41–43 dBA under a heavier dense-70B load (~147–160W). At idle they’re within 1 dBA of each other. The EVO-X2 wins idle power instead: 8–14W versus 16–25W.

Can I upgrade the RAM later on either machine? No. The 128GB LPDDR5X is soldered on both. Storage is upgradeable on both — two M.2 2280 PCIe 4.0 slots, up to 16TB total.

Should I wait for prices to drop before buying? The DRAM shortage has moved both boxes hundreds of dollars in both directions since launch — the EVO-X2 peaked near $3,200 in late 2025 and settled near $2,199 by mid-2026. There’s no announced event that resets memory pricing, so buy the cheaper of the two when you actually need the machine rather than timing the market.

Sources

Last updated September 12, 2026. Prices and specs change; verify current rates before purchasing.

Was this article helpful?