Minisforum MS-S1 Max vs GMKtec EVO-X2 in 2026: The $800 Premium the DRAM Crisis Just Erased

By RunAIHome Team · 14 min read
mini-pcstrix-haloryzen-ai-maxminisforumgmkteclocal-llmhardwarecomparison

TL;DR: Both machines run the same Ryzen AI Max+ 395 with 128GB of LPDDR5X-8000 at 256 GB/s, so LLM decode speed is a tie at roughly 31 tok/s on gpt-oss 120B. The MS-S1 Max launched $800 over the EVO-X2’s street price to pay for dual 10GbE, two 80Gbps USB4 V2 ports, a PCIe slot, and a built-in 320W PSU. In September 2026 the DRAM crisis inverted that gap: the better-equipped box is now the cheaper one.

Minisforum MS-S1 MaxGMKtec EVO-X2Used RTX 3090 build
Best for128GB inference plus a real home-lab I/O board128GB inference if its price dips back downEverything that fits in 24GB
Price (Sep 2026)~$2,959 street (Amazon/Newegg), $2,299 at launch$3,499 direct (128GB/1TB); was $2,199–$2,299 in June~$1,070 card, ~$1,400 build
gpt-oss 120B~30–35 tok/s~31 tok/swon’t fit
Expansiondual 10GbE, 2× USB4 V2 (80Gbps), PCIe 4.0 x16 slot (x4 electrical)dual 2.5GbE, 2× USB4 (40Gbps), no slotfull x16 slot, your case, your rules
The catchPCIe slot is x4 and cramped; second M.2 is x1price spiked ~60% in six monthshard 24GB ceiling

Honest take: Same chip, same tokens per second — but this month the Minisforum MS-S1 Max is both the better-equipped machine and the cheaper one, which makes it the first Strix Halo recommendation this site has made that doesn’t need a tiebreaker. Buy the EVO-X2 only if you find it meaningfully under $2,959 the week you order.

The queue for this comparison was written when the question was “what does the MS-S1 Max’s $800 premium actually buy?” Six months of memory repricing later, the honest question has become “why is the box with more I/O now cheaper than the box with less?” Both versions have the same answer: the premium was never about speed, and neither is the discount.

Token speed: settled before either box was designed

If you’ve read our Beelink GTR9 Pro vs EVO-X2 comparison from this morning, you already know how this section ends. Every 128GB Strix Halo machine — Minisforum, GMKtec, Beelink, Framework — pairs the same 16-core Ryzen AI Max+ 395 and Radeon 8060S iGPU with the same 128GB of LPDDR5X-8000 on a 256-bit bus: 256 GB/s of memory bandwidth. Token generation is bandwidth-bound, so the decode numbers are interchangeable.

Measured, not assumed: AkitaOnRails ran gpt-oss 120B in its native MXFP4 on the MS-S1 Max under llama.cpp (Vulkan, after a driver update) and landed at 30–35 tok/s at small context. The EVO-X2 posts ~31 tok/s on the same model, and dense Llama 3.3 70B decodes at ~5 tok/s on either machine — the numbers from our EVO-X2 review transfer one-for-one. Smaller models behave the same way: 7B–13B classes in the 30–45 tok/s range, 30B-class MoE models at 70–100 tok/s.

One asymmetry worth knowing: Minisforum ships four power modes from 110W up to a 160W sustained ceiling — the highest TDP configuration any vendor has shipped this chip at. That extra headroom does nothing for decode (bandwidth-bound), but prompt processing is compute-bound, so the 160W mode is the one place the MS-S1 Max can pull ahead on long-context work. Treat it as a margin, not a tier: nobody has published a controlled prefill shootout between the two, and we won’t invent one.

So the real comparison is everything around the chip. That’s where these boxes stop being twins.

What the launch premium actually bought

At launch the MS-S1 Max cost $2,299 (128GB/2TB) against the EVO-X2’s $1,999–$2,199. Line the spec sheets up and the $800-ish delta was itemizable:

MS-S1 MaxEVO-X2
Wired networking2× 10GbE (Realtek RTL8127)2× 2.5GbE (Realtek 8125BG)
USB42× USB4 V2 80Gbps + 2× USB4 40Gbps2× USB4 40Gbps
PCIe expansionx16 slot, PCIe 4.0 x4 electricalnone
Power supplyinternal 320W, no brickexternal adapter
M.2 storage2 slots (x4 + x1), up to 8TB, RAID 0/12 slots (both usable PCIe 4.0), up to 16TB
APU power modes110–160W~120–140W typical

Three of those rows matter for a home AI lab.

The networking is the same argument as the GTR9 Pro, minus the drama. Dual 10GbE moves a 40GB GGUF off your NAS in under a minute instead of ~2.5 minutes over 2.5GbE, every time you swap models. And the controller is worth naming: the RTL8127 is the same Realtek part Beelink switched to when it revised the GTR9 Pro away from its problematic Intel E610 boards. Minisforum shipped the boring-and-working option from day one. Minisforum’s own marketing pushes this further — a two-unit cluster over 10GbE running Qwen3 235B Q4 at 10.87 tok/s — but that’s a vendor demo, not an independent benchmark, and distributed inference adds capacity, not speed. Buy the second box for the model that doesn’t fit in 128GB, never for throughput.

USB4 V2 is the quiet future-proofing. Minisforum put a discrete controller on the board to drive two 80Gbps ports — Thunderbolt 5-class bandwidth for external NVMe arrays or an eGPU dock, on an AMD platform. The EVO-X2’s 40Gbps ports are fine today; they’re just half the ceiling.

The PCIe slot is real, with real fine print. More on that next, because it’s the most misunderstood spec on this machine.

The PCIe slot: an I/O slot, not a GPU bay

The slot is physically x16 but electrically PCIe 4.0 x4 — about 8 GB/s of bandwidth. NASCompares’ teardown adds the practical constraints: clearance pushes you toward half-height, half-length cards, the board area gets little direct airflow (the main fans serve the APU on the other side of the chassis), and the shared 320W internal PSU has no supplemental GPU power to spare.

So what is it for? Exactly what Minisforum’s MS-01 workstation line taught people to use it for: a 25GbE NIC, a quad-M.2 carrier for model storage, a capture card, a compact accelerator. If you drop a card in, verify what link it actually trained to:

sudo lspci -vv | grep -A1 'LnkCap\|LnkSta'
# LnkCap: Port #0, Speed 16GT/s, Width x4   <- slot ceiling
# LnkSta: Speed 16GT/s, Width x4            <- card negotiated x4: expected

If you were hoping to slide a used RTX 3090 in and get the best of both worlds — 128GB of slow unified memory and 24GB at 936 GB/s — the chassis says no before the electrical spec does. A blower-style single-slot card like an RTX A4000 physically fits some builds, but at x4 with no PSU headroom you’re assembling a compromise, not a workstation. The honest read: the slot raises the machine’s ceiling as a server (networking, storage), not as a GPU host. For an actual dGPU tower, a used RTX 3090 build remains the answer for everything under 24GB.

The storage row cuts the other way, and it’s the MS-S1 Max’s one genuine spec loss: its second M.2 slot runs at PCIe 4.0 x1 (~2 GB/s), and total capacity tops out at 8TB versus the EVO-X2’s two full-speed slots and 16TB ceiling. A 60GB model file loads noticeably slower from the x1 slot. Fix is cheap — put your model library on the x4 slot’s drive and your OS on the x1 — but know which slot is which before you image the machine.

Noise and power: pick your mode, mind the idle

The MS-S1 Max’s four power modes make single-number noise claims useless, so here are Notebookcheck’s measurements per mode: Performance 37 dB idle / 48 dB load; Balanced 35 / 42; Quiet 32 / 41.6. ServeTheHome measured about 43 dBA at one meter under sustained AI load in performance mode, with the wall meter reading ~120W during gpt-oss 120B inference and 9–12W idle. Notebookcheck’s complaint — the machine is louder than it should be at idle in performance mode — has a two-keystroke solution: if the box is a 24/7 inference server, run Balanced or Quiet; you lose sustained CPU wattage that decode (bandwidth-bound, remember) barely uses, and you drop 5 dB at idle.

The EVO-X2’s corresponding numbers, from Tom’s Hardware and ServeTheHome testing: 8–14W idle, 147–160W wall under dense-70B load, 37–38 dBA idle and 41–43 dBA under that load. Under comparable AI loads the two boxes are within a couple of dBA of each other, and both idle in the same 8–14W band once the Minisforum is out of performance mode. At $0.12/kWh the idle-power difference rounds to a few dollars a year — the 24/7 power math doesn’t pick a winner here.

The price inversion: what six months of DRAM crisis did

This is the section the queue premise didn’t see coming. The trajectory, with dates:

  • MS-S1 Max (128GB/2TB): pre-orders at $2,099, launch at $2,299 (VideoCardz called it a “premium price tag” at the time). Through the memory crunch the Amazon listing peaked at $3,299.99, and as of September 2026 sits at $2,959 — matching the $2,959 Newegg price TechRadar flagged as a deal against an inflated $5,699 list.
  • GMKtec EVO-X2 (128GB): launched near $1,999, sat at $2,199 direct / $2,299 at Micro Center in June 2026 — the numbers still in our morning GTR9 Pro comparison — then repriced hard: GMKtec’s US store now lists a 128GB/1TB configuration at $3,499, and price tracking shows the 128GB config climbing from $2,099 to $3,299 in six months, a ~60% increase. The $1,499 figure that still circulates is the 64GB/1TB config; no 128GB unit ever sold for that.

Read the September snapshot carefully: the MS-S1 Max at $2,959 ships 2TB of storage, dual 10GbE, and 80Gbps USB4 V2; the EVO-X2 at $3,499 ships 1TB and 2.5GbE. The machine that justified a premium on spec now undercuts the machine it out-specs by ~$540. That’s not Minisforum getting cheaper — it’s soldered LPDDR5X repricing at different speeds through two vendors’ inventory. Which means the inversion can un-invert: if a restocked EVO-X2 batch lands under the Minisforum’s street price, the value crown moves back. Check both listings the week you order. What doesn’t move is the spec sheet — at equal prices, the MS-S1 Max wins on I/O every time.

Who should buy which

Buy the MS-S1 Max if you’re buying this class of machine at all right now. It’s the cheaper box this month, and the extras aren’t decorative for a home lab: 10GbE to your NAS, a slot for the 25GbE card or M.2 carrier you’ll want later, USB4 V2 for external model storage, and a 160W mode for long-context prefill. Set it to Balanced mode if it lives in your office. If the box will anchor a local coding stack, the software side — routing Claude Code or other agents to local models — is its own project.

Buy the EVO-X2 if its price falls back under the Minisforum’s. Nothing is wrong with it — same speed, proven thermals, 16TB storage ceiling, and a longer track record shipping this platform than anyone. It’s simply not worth a $540 premium over a machine that out-specs it. At $2,199 it was the segment default; at $3,499 it isn’t.

Buy neither if your models fit in 24GB. The rule from every Strix Halo article on this site holds: these boxes exist to hold models a consumer GPU can’t. Run your actual model-plus-context through the VRAM calculator first; if the answer is 24GB or less, a used RTX 3090’s 936 GB/s decodes roughly 3× faster for a third of the money. Still unsure 128GB is your workload? Rent one evening on RunPod and find out before spending $3,000. The premium alternatives — DGX Spark and AMD’s dev kit — are covered in our 128GB head-to-head, and the cheaper end of the segment in the mini PC guide.

What to actually buy

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

Your situationThe machinePriceWhere
100B+ MoE models at home, on a real home-lab networkMinisforum MS-S1 Max (128GB/2TB)~$2,959Check price
Same workload, and the price dipped below the Minisforum this weekGMKtec EVO-X2 (128GB)~$3,499 direct, watch for dipsCheck 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 Minisforum MS-S1 Max faster than the GMKtec EVO-X2 for local LLMs? Not where it counts. Both decode gpt-oss 120B at ~31 tok/s because both have identical 256 GB/s memory systems. The MS-S1 Max’s 160W power ceiling can help compute-bound prompt processing on long contexts, but token generation speed is a tie.

Can I install a graphics card in the MS-S1 Max’s PCIe slot? Treat the slot as I/O expansion, not a GPU bay. It’s x16 physical but PCIe 4.0 x4 electrical, clearance favors half-height cards, the area gets little airflow, and the internal 320W PSU has no supplemental GPU power. It’s excellent for a 25GbE NIC or M.2 storage carrier.

Why is the EVO-X2 suddenly more expensive than the MS-S1 Max? Soldered LPDDR5X repriced through the 2026 DRAM shortage at different speeds for different vendors. The EVO-X2’s 128GB config climbed from $2,099 to $3,299–$3,499 in about six months, while the MS-S1 Max’s street price settled near $2,959. The gap can flip back — check both before ordering.

Can I upgrade the RAM on either machine? No. The 128GB LPDDR5X is soldered on both. Storage differs: the EVO-X2 takes two full-speed M.2 drives up to 16TB; the MS-S1 Max’s second M.2 slot is only PCIe 4.0 x1 and total capacity caps at 8TB — put your model library on its x4 slot.

Is the MS-S1 Max loud? In performance mode it idles at 37 dB, which reviewers dinged. Balanced mode (35 dB idle / 42 dB load) or Quiet mode (32 / 41.6) fixes it, and decode speed barely changes because token generation doesn’t use the extra CPU wattage anyway.

Sources

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

Was this article helpful?