Six months ago, an AMD Ryzen AI Max+ 395 mini PC with 128GB of unified memory cost around $1,500 to $2,200 depending on the brand. In September 2026, the same hardware routinely lists for $3,500 to $4,349. Nothing changed inside the chip. What changed is the price of DRAM, and it has turned a category of machine built to make local AI affordable into one of the most confusing purchases in PC hardware. Three boxes now dominate the conversation among developers who want to run large language models without a cloud bill: the GMKtec EVO-X2, the Beelink GTR9 Pro, and the Minisforum MS-S1 Max. All three use the same AMD “Strix Halo” silicon. None of them cost the same amount, ship the same memory tiers, or make the same tradeoffs on networking and storage. This comparison works through the actual specs, the benchmark numbers that exist in public reviews, current September 2026 pricing, and which one of the three is worth buying depending on what you actually plan to run on it.
Why three Strix Halo mini PCs are suddenly worth comparing
A year ago, running a 70-billion-parameter language model at home meant either renting a cloud GPU by the hour or building a multi-GPU tower with enough VRAM to hold the weights. AMD’s Ryzen AI Max+ 395 changed the math by putting a 16-core Zen 5 CPU, a 40-compute-unit RDNA 3.5 GPU, and a 50-TOPS XDNA 2 neural processing unit on one die, then pairing it with up to 128GB of LPDDR5X-8000 memory that both the CPU and GPU can share. That unified pool is the entire pitch: a discrete GPU with 128GB of VRAM does not exist at any consumer price point, but a Strix Halo mini PC gets close enough to run large quantized models that would otherwise need a multi-thousand-dollar workstation card.
What makes September 2026 a strange time to buy one is the memory market itself. DRAM and NAND flash pricing has climbed sharply through the year, and it has hit machines that are almost entirely memory by design harder than most. Valve raised Steam Deck OLED pricing by roughly 44 to 46 percent in May 2026, citing memory and storage costs, and the effect has been even sharper on 128GB Strix Halo boxes, where memory is most of the bill of materials. Tom’s Hardware reviewed the GMKtec EVO-X2 in February 2026 at $1,499 for its test configuration. By late September, comparable listings for the same machine had more than doubled. That volatility is the backdrop for every price comparison in this article, and it is the single biggest reason to check current listings before buying rather than trusting any number more than a few weeks old.
What the Ryzen AI Max+ 395 actually does differently
All three machines in this comparison run the identical processor, which is unusual for a three-way hardware comparison and makes this fundamentally a comparison of memory tiers, ports, cooling, and price rather than raw compute. Understanding what the chip provides matters before comparing the boxes built around it.
The XDNA 2 NPU and 50 TOPS
The Ryzen AI Max+ 395 packs a dedicated XDNA 2 neural processing unit rated at 50 TOPS (trillions of operations per second) for INT8 workloads. In practice, the NPU handles lightweight, always-on AI tasks such as webcam background processing, voice transcription, and some Windows Copilot+ features, while the heavy lifting for large language model inference happens on the integrated GPU, not the NPU. Anyone buying one of these three machines specifically to run 30B-plus parameter models should think of the NPU as a bonus for efficiency-focused background tasks, not the component that determines whether a given model will run.
Unified memory, explained without the marketing
The Radeon 8060S GPU built into the Ryzen AI Max+ 395 has 40 RDNA 3.5 compute units and no dedicated VRAM of its own. Instead, it draws from the same LPDDR5X-8000 pool as the CPU, with roughly 256GB/s of theoretical memory bandwidth shared between the two. On the 128GB configuration, reviewers have found that up to 96GB can be allocated to the GPU for model weights, leaving the rest for the operating system and CPU tasks. That bandwidth figure is the real ceiling on performance: it is roughly a seventh of what a high-end discrete GPU like an RTX 5090 offers, which is why these machines run large models but not at the speeds a $2,000 graphics card delivers on smaller ones. The tradeoff is capacity for speed, and it is a genuinely useful tradeoff for models too large to fit on any single consumer GPU.
Full specs comparison: EVO-X2 vs GTR9 Pro vs MS-S1 Max
Here is how the three machines stack up on paper. All three ship with the same processor and GPU, so the differentiation comes almost entirely from memory tiers, storage, ports, and price.
| Spec | GMKtec EVO-X2 | Beelink GTR9 Pro | Minisforum MS-S1 Max |
|---|---|---|---|
| CPU | Ryzen AI Max+ 395, 16C/32T | Ryzen AI Max+ 395, 16C/32T | Ryzen AI Max+ 395, 16C/32T |
| Boost clock | Up to 5.1GHz | Up to 5.1GHz | Up to 5.1GHz |
| NPU | XDNA 2, 50 TOPS | XDNA 2, 50 TOPS | XDNA 2, 50 TOPS |
| GPU | Radeon 8060S, 40 CU RDNA 3.5 | Radeon 8060S, 40 CU RDNA 3.5 | Radeon 8060S, 40 CU RDNA 3.5 |
| RAM tiers confirmed | 64GB / 96GB / 128GB LPDDR5X-8000 | 128GB LPDDR5X-8000 (only tier confirmed) | 64GB / 128GB LPDDR5X-8000 |
| Memory bandwidth | ~256GB/s | ~256GB/s | ~256GB/s |
| Storage | 1 M.2 2280 slot, 1-2TB shipped, up to two 8TB drives | Dual M.2 2280 slots, 2TB shipped, up to 16TB total | Multiple M.2 NVMe slots (workstation chassis) |
| Networking | 2.5GbE, Wi-Fi 7 | Dual 10GbE, Wi-Fi 7 | High-speed Ethernet, Wi-Fi |
| USB | Dual USB4, USB-A | Dual USB4 (40Gbps), USB-A | USB4-class ports |
| Display out | HDMI, DisplayPort | HDMI 2.1, DisplayPort 2.1 | HDMI, DisplayPort |
| Rated power draw | ~120W sustained, 140W peak | ~125-128W under LLM load | Not independently confirmed |
| First independent review | Tom’s Hardware, Feb 2026 | Specialist 2026 review coverage | TechRadar, 2026 |
| 128GB September 2026 price | ~$3,499.99-$3,649.99 | ~$4,349 | ~$3,599-$3,799 |
The pattern that emerges is straightforward. The EVO-X2 offers the broadest range of memory tiers, which matters if 128GB is more capacity than you actually need. The GTR9 Pro skips smaller tiers entirely and instead pours its engineering budget into dual 10-gigabit Ethernet and dual M.2 slots, positioning it as much as a small server as a desktop. The MS-S1 Max sits in between on price while offering a workstation-style chassis with room for more drives than the EVO-X2.
GMKtec EVO-X2: the value play with the most independent testing
The EVO-X2 has the longest public review trail of the three, including a Tom’s Hardware review published in February 2026, additional testing from ServeTheHome, and a separate write-up at PCWorld. That review coverage matters because it means GMKtec’s numbers are the most independently checked of the three, rather than resting primarily on manufacturer claims.
The machine offers three memory tiers: 64GB, 96GB, and 128GB, all LPDDR5X-8000. That range is the EVO-X2’s biggest practical advantage over the GTR9 Pro, which has only been confirmed in a 128GB configuration. A developer who mainly wants to run 30B-class models with headroom, rather than 70B-plus models, can buy the 64GB tier for meaningfully less money rather than paying for capacity they will not use. Storage is a single M.2 2280 slot in most configurations, shipping with 1TB or 2TB, with GMKtec’s spec sheet claiming support for up to two 8TB drives in higher-end SKUs. Networking is more modest than the GTR9 Pro, with 2.5-gigabit Ethernet and Wi-Fi 7 rather than the Beelink’s dual 10-gigabit ports.
Pricing has been the EVO-X2’s most volatile trait. Tom’s Hardware’s review unit came in at $1,499, and a 96GB/2TB configuration was listed at $1,799.99 around the same period. By September 2026, comparable listings showed 64GB/1TB models around $1,999.99 to $2,199.99, and 128GB models between $3,499.99 and $3,649.99 depending on storage. That is not a small drift. It roughly doubles the entry price of the machine within about seven months, and it tracks closely with the same memory-pricing pressure covered in tech-insider.org’s report on DRAM prices forcing AYANEO to halt handheld production earlier in 2026.
Beelink GTR9 Pro: built for networking and server duty
The GTR9 Pro is the outlier of the three in both connectivity and price. It is only confirmed to ship in a 128GB LPDDR5X-8000 configuration, so there is no cheaper entry point the way there is with the EVO-X2 or MS-S1 Max. What it does offer in exchange is the strongest I/O of the group: dual Intel 10-gigabit Ethernet ports, dual USB4 ports rated at 40Gbps, HDMI 2.1, DisplayPort 2.1, Wi-Fi 7, and dual M.2 2280 NVMe slots that Beelink rates for up to 16TB of combined storage. It also ships with an integrated power supply rather than a bulky external brick, which matters if the machine is going into a rack or a cramped homelab shelf rather than sitting on a desk.
That connectivity positions the GTR9 Pro less as a desktop workstation and more as a compact inference server that happens to also work as a desktop. Dual 10GbE is overkill for casual use but genuinely useful if the box is serving an LLM API endpoint to other machines on a home or office network, or acting as a NAS-plus-inference box in a small studio. Beelink’s own testing put the GTR9 Pro at 31.41 tokens per second on GPT-OSS 120B, a large mixture-of-experts model, while drawing roughly 125 to 128 watts, and around 5 tokens per second on a dense 70B model — numbers broadly in line with what the same silicon achieves in the EVO-X2, since the underlying GPU and memory bandwidth are identical.
The catch is price. Late-September 2026 listings put the 128GB GTR9 Pro at roughly $4,349, which is $700 to $850 more than comparable 128GB configurations of the EVO-X2 or MS-S1 Max. That premium buys the networking and dual-NVMe advantages described above, but it is a real premium, not a rounding error, and it only makes sense for buyers who will actually use the extra Ethernet bandwidth and drive bays.
Minisforum MS-S1 Max: the workstation alternative
Minisforum’s entry sits in a workstation-style chassis with room for multiple NVMe drives, and it is worth being precise about naming here: it is the MS-S1 Max, not the older MS-S1 without the Max suffix, and not Minisforum’s other Ryzen-based systems like the AtomMan G1 Pro, which use different chips entirely. The MS-S1 Max is confirmed in 64GB and 128GB LPDDR5X-8000 configurations, with a 96GB tier not established in current listings the way it is for the EVO-X2.
TechRadar’s review of the MS-S1 Max found meaningful promotional pricing on the machine earlier in 2026, with direct pricing around $2,499, down from a $2,999 list price, before the memory market tightened further through the year. By September 2026, 128GB configurations were listed between roughly $3,599.99 and $3,799, which places the MS-S1 Max below the GTR9 Pro but above the EVO-X2’s 128GB pricing in most listings. Port and connectivity details for the MS-S1 Max have been less consistently documented across retailers than the other two machines, though multiple NVMe slots and USB4-class connectivity are confirmed. Buyers considering this machine should check the specific retailer listing closely, since Minisforum’s product pages have historically varied more between regional SKUs than GMKtec’s or Beelink’s.
The MS-S1 Max’s strongest case is as a Minisforum-ecosystem alternative for buyers already using the company’s other mini PCs or docking accessories, and as a slightly cheaper path to 128GB than the GTR9 Pro without giving up the workstation-style chassis and multiple drive bays that the single-slot EVO-X2 lacks.
September 2026 pricing across every configuration, and why it keeps moving
Pricing on all three machines has been unusually unstable through 2026, and it is worth showing the range rather than a single snapshot number, since a figure captured in August looks stale within weeks.
| Machine / config | Early-to-mid 2026 price | September 2026 price | Approx. change |
|---|---|---|---|
| EVO-X2 64GB/1TB | Not directly comparable tier at launch | ~$1,999.99-$2,199.99 | Entry tier, newly priced |
| EVO-X2 review config (Feb 2026) | $1,499 | N/A (superseded) | Baseline for comparison |
| EVO-X2 128GB/1TB | ~$2,199 (Feb 2026 est.) | ~$3,499.99 | Roughly +59% |
| EVO-X2 128GB/2TB | ~$1,799.99 (96GB/2TB, Feb 2026) | ~$3,649.99 | Steep run-up |
| GTR9 Pro 128GB | Not confirmed at a lower historical price | ~$4,349 | Highest of the three |
| MS-S1 Max 128GB | ~$2,499 (promo, earlier 2026) | ~$3,599-$3,799 | Roughly +44% to +52% |
| Steam Deck OLED 512GB (reference) | $549 (original) | $789 | +43.7% |
| Steam Deck OLED 1TB (reference) | $649 (original) | $949 | +46.2% |
The Steam Deck OLED rows are included deliberately as a reference point, not because it competes with these mini PCs, but because Valve’s confirmed 44 to 46 percent price increase in May 2026 gives a documented, company-stated percentage for how much memory and storage costs rose industry-wide during the year. The Strix Halo mini PC market, which is far more memory-dependent per unit than a handheld, has moved by a similar or larger percentage depending on the specific configuration. That mirrors what tech-insider.org covered when Acer warned that PC prices broadly were climbing 20 percent due to the memory crisis, and the effect compounds further in machines like these where 128GB of premium LPDDR5X is a large fraction of the total bill of materials. It is also consistent with reporting that DRAM prices jumped by a record margin through 2026, pushing relief for memory-heavy hardware categories out to 2028 in some forecasts.
The practical takeaway: treat every price in this article as a snapshot, not a guarantee. Check current retailer listings directly before ordering, and if a price looks unusually low compared to the ranges above, verify it is not a discontinued or out-of-stock configuration.
It is also worth separating manufacturer list price from what buyers actually pay. All three brands sell direct as well as through Amazon, Newegg, and regional resellers, and discount timing has been inconsistent through 2026 as each company adjusts to component costs at different points in their supply chain. GMKtec’s direct store has shown the most frequent short-term promotions of the three, occasionally undercutting the ranges shown above for a matter of days before reverting. Beelink and Minisforum have both run region-specific coupon codes that can shift the GTR9 Pro or MS-S1 Max price by several hundred dollars depending on where the buyer is located. None of that changes the overall ordering of which machine costs more than which, but it means the exact figure at checkout can land meaningfully below the listing price quoted in a given review if the timing is right.
Benchmarks: Geekbench, 3DMark, and tokens per second
Because all three machines share the identical Ryzen AI Max+ 395 and Radeon 8060S combination, raw compute benchmarks land in a similar range across all three, and the differences that show up in independent testing are closer to run-to-run variance and cooling design than to any real architectural gap. The most detailed public numbers come from the EVO-X2, since it has the deepest independent review coverage.
| Benchmark | GMKtec EVO-X2 | Beelink GTR9 Pro | Source |
|---|---|---|---|
| Geekbench 6 single-core | ~3,076 | Not independently published | Independent reviewer testing |
| Geekbench 6 multi-core | ~20,000 | Not independently published | Independent reviewer testing |
| 3DMark Time Spy | ~10,554-11,253 | Not independently published | Independent reviewer testing |
| 7B-13B quantized LLM (tok/s) | ~30-45 | Similar range expected (same silicon) | Independent local-AI testing |
| Sparse 30B MoE model (tok/s) | ~70-100 | Similar range expected | Independent local-AI testing |
| Dense 70B model, heavy quant (tok/s) | ~5 | ~5 | Independent local-AI testing on both |
| GPT-OSS 120B MoE (tok/s) | Not separately published | 31.41 at ~125-128W | Beelink review testing |
| Qwen3-235B-A22B MoE (tok/s) | ~11 (128GB unit) | Not separately published | Independent local-AI testing |
Two things stand out. First, dense models scale poorly on this hardware once they cross roughly 70 billion parameters, because the bottleneck is memory bandwidth, not memory capacity, and a dense 70B model has to move every one of its weights through that ~256GB/s pipe for every token generated. Second, mixture-of-experts models are a far better fit for this architecture, since only a fraction of the total parameters activate per token. That is why GPT-OSS 120B, a MoE model, runs at 31 tokens per second on the GTR9 Pro while a dense 70B model on the same hardware crawls at roughly 5 tokens per second, despite the MoE model technically having more total parameters. Anyone buying one of these machines specifically for local inference should prioritize MoE-architecture models over dense ones of similar or even larger nominal size.
How much AI model each RAM tier can actually run
Model capacity depends heavily on quantization level, context length, and how much memory the operating system and inference runtime reserve for themselves, but the following ranges reflect what independent local-AI testers have found practical, not just theoretically possible, across Strix Halo machines.
| Usable memory | Practical model range | Notes |
|---|---|---|
| 32GB | 7B-14B comfortably; some 20B-30B with compromises | Not a tier offered by any of these three machines, shown for reference |
| 64GB (EVO-X2, MS-S1 Max) | 30B-34B comfortably; 70B possible only with aggressive quantization and reduced speed | Cheapest entry point to Strix Halo capacity |
| 96GB (EVO-X2) | 70B-class models become practical | Middle tier, EVO-X2-only among these three |
| 128GB (all three) | 70B dense models run at reduced speed; 100B-120B MoE models run at usable speeds | GPT-OSS 120B and Qwen3-235B-A22B are realistic targets at this tier |
The practical guidance follows directly from this table: buyers who mainly want to run 7B to 30B parameter models for coding assistance, drafting, or light agentic workflows do not need the 128GB tier at all, and should strongly consider the EVO-X2’s 64GB configuration to avoid paying the current DRAM-crisis premium for capacity they will not use. Buyers who specifically want to run 70B-plus dense models or large MoE models like GPT-OSS 120B need the full 128GB, and should expect single-digit to low-double-digit tokens per second rather than the fast responses a cloud API delivers.
How these compare to Nvidia DGX Spark and Apple Mac Studio
None of these three Strix Halo machines exist in a vacuum. The two other well-known paths to a large local unified-memory pool are Nvidia’s DGX Spark and Apple’s Mac Studio, and the choice between architectures matters as much as the choice between the three AMD boxes covered above.
DGX Spark tops out at 128GB of unified memory, roughly matching the ceiling on these Strix Halo machines, but its real advantage is software: it runs on Nvidia’s CUDA stack, which remains the most mature and widely supported ecosystem for AI inference and training tooling. Buyers whose workflow depends on CUDA-only libraries, or who want guaranteed compatibility with Nvidia-optimized inference engines, will get a smoother experience on DGX Spark even at similar memory capacity, as detailed in tech-insider.org’s comparison of DGX Spark against the Mac Studio’s up-to-512GB memory ceiling.
Mac Studio goes further on raw capacity, with configurations reaching up to 512GB of unified memory on Apple’s high-end silicon, which opens the door to running models that no Strix Halo or DGX Spark machine can hold at all. Its ecosystem is Metal and MLX rather than CUDA or ROCm, which works well for Apple-optimized tools but less well for Linux-first or CUDA-dependent AI stacks. AMD’s Strix Halo machines sit as the value option of the three: none of the GMKtec, Beelink, or Minisforum machines match Mac Studio’s memory ceiling or Nvidia’s software maturity, but all three undercut both on price per gigabyte of unified memory, at least before accounting for the DRAM-driven price increases that have narrowed that gap significantly since early 2026. AMD’s own roadmap also points toward the issue continuing to evolve: the enterprise-focused Ryzen AI Max PRO 400, covered separately by tech-insider.org with up to 192GB of unified memory aimed at running 300B-parameter models, signals where this category is headed next, even though it is not yet shipping in consumer mini PCs like the three compared here.
Noise, thermals, and running one of these 24/7
A detail that gets less attention than raw benchmark numbers, but matters enormously if one of these machines is going to sit on a desk or in a closet generating tokens around the clock, is how loud and hot it gets under sustained load. The Ryzen AI Max+ 395’s 120W to 140W power envelope is modest next to a discrete GPU workstation, but it is still concentrated in a chassis far smaller than a mid-tower desktop, and all three vendors handle that heat differently.
The EVO-X2’s smaller single-fan cooling solution is adequate for burst workloads like a coding assistant answering occasional prompts, but independent testing has noted fan noise climbing audibly during sustained LLM inference or back-to-back batch image generation, and thermal throttling has been observed on some units after extended periods at full load. The GTR9 Pro’s larger chassis, built partly to accommodate its dual-Ethernet and dual-NVMe hardware, gives it more room for airflow and a quieter sustained-load profile in the reviews that have covered it, which lines up with its positioning as a box meant to run unattended as a server rather than sit next to someone at a desk. The MS-S1 Max’s workstation-style case similarly trades desktop compactness for better thermal headroom, though independent acoustic testing for it is thinner than for the other two.
The practical implication: anyone planning to leave one of these machines running inference jobs overnight or hosting an always-on API endpoint should weight the GTR9 Pro or MS-S1 Max more heavily than someone who only spins up a model for a few minutes at a time during the workday, where the EVO-X2’s more compact cooling is much less likely to be a problem. Placement matters too. All three benefit from open airflow rather than being tucked into an enclosed cabinet, and none of them are quiet enough to disappear into a bedroom setup the way a passively cooled NAS might.
5 real-world use cases: who should actually buy one
Specs and benchmarks only matter in the context of what someone is actually going to do with the machine. Here are five scenarios where a Strix Halo mini PC is a genuinely good fit, and which of the three machines suits each one best.
1. A self-hosted coding and writing assistant. Developers running 14B-32B coding models through tools like Continue, Cline, or a local Ollama endpoint get fast enough responses on any of these three machines, and the 64GB EVO-X2 is the cheapest entry point that still leaves comfortable headroom.
2. A small studio or agency running local image and video AI. Studios that want to avoid per-generation cloud GPU rental fees for Stable Diffusion or local video models benefit from the GPU compute here, though the memory bandwidth ceiling means this is a better fit for batch jobs run overnight than for real-time interactive generation.
3. A homelab AI inference server serving multiple users or devices. This is the Beelink GTR9 Pro’s clearest use case: dual 10GbE and dual USB4 ports make it well suited to sitting in a rack or closet and serving an OpenAI-compatible API endpoint to other machines on the network, something the single-NIC EVO-X2 and MS-S1 Max are not built around as directly.
4. A small business handling sensitive documents that cannot go to a cloud API. Law firms, healthcare practices, and financial advisors experimenting with document summarization or retrieval-augmented generation under compliance constraints get a genuine on-premises option here, particularly on the 96GB or 128GB tiers where larger, more accurate models fit.
5. A Proxmox or homelab virtualization host with an AI sidecar. The MS-S1 Max’s workstation-style chassis and multiple NVMe slots make it a reasonable pick for someone who wants to run several VMs or containers alongside an occasional local-AI workload, rather than dedicating the whole machine to inference around the clock.
Migration guide: moving from a cloud AI API to a local Strix Halo box
Switching a workflow that currently calls a cloud AI API over to a local Strix Halo mini PC is a software exercise more than a hardware one, since the OpenAI-compatible API format has become close to a de facto standard for local inference servers. The steps below apply to any of the three machines covered in this comparison.
Step 1: Decide your target model size before you buy. Use the RAM-tier table above. If your workload tops out around 30B parameters, buy 64GB and save money. If you need 70B-plus or large MoE models, budget for 128GB.
Step 2: Choose an OS. All three machines ship with Windows 11, but Linux distributions such as Ubuntu or Fedora generally offer more mature AMD ROCm support for GPU-accelerated inference as of September 2026.
Step 3: Install an inference runtime. Ollama and llama.cpp both support Vulkan and ROCm backends that take advantage of the Radeon 8060S’s compute units, and both expose OpenAI-compatible REST endpoints out of the box, as previously detailed in tech-insider.org’s guide to running an LLM locally and its llama.cpp setup walkthrough.
# Install llama.cpp with Vulkan backend for Radeon 8060S
git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
cmake -B build -DGGML_VULKAN=ON
cmake --build build --config Release
# Download a quantized GGUF model, then serve an OpenAI-compatible endpoint
./build/bin/llama-server -m ./models/model-name-Q4_K_M.gguf
--host 0.0.0.0 --port 8080 -ngl 99 -c 8192Step 4: Download quantized models sized to your RAM tier. Q4_K_M or Q5_K_M quantizations from Hugging Face give a reasonable balance of quality and memory footprint. A similar approach was covered in tech-insider.org’s guide to running Qwen3.8-27B locally with Ollama.
Step 5: Point your existing application at the local endpoint. Most tools that call the OpenAI API accept a custom base URL. Changing that one setting to point at your mini PC’s local IP address and port is usually the only code change required, since the request and response format stays identical.
Step 6: Benchmark before committing fully. Run representative prompts through the local model and compare quality and latency against the cloud API you are replacing before cutting over production workflows, since local inference at these token-per-second rates will feel noticeably slower for large models than a cloud API call.
Pros and cons of each machine
GMKtec EVO-X2 pros: broadest range of memory tiers, deepest independent review coverage, lowest entry price of the three. EVO-X2 cons: weakest networking of the group at 2.5GbE, single primary M.2 slot limits storage expansion compared to the other two.
Beelink GTR9 Pro pros: dual 10GbE and dual USB4 make it the strongest networked-server option, dual M.2 slots support up to 16TB of storage, integrated power supply avoids a bulky external brick. GTR9 Pro cons: only ships in a 128GB configuration so there is no cheaper entry point, and it carries the highest price of the three at roughly $4,349.
Minisforum MS-S1 Max pros: workstation-style chassis with room for multiple drives, historically available at aggressive promotional pricing, sits between the other two on 128GB pricing. MS-S1 Max cons: port and connectivity details are less consistently documented across retailers than the other two, and no 96GB tier is confirmed for buyers who want a middle option.
The verdict: which Strix Halo mini PC should you buy
For most buyers, the GMKtec EVO-X2 is the sensible default. It has the most independent review data backing its claimed performance, it is the only one of the three offering a genuinely cheaper 64GB entry point, and even at inflated September 2026 pricing it typically undercuts the GTR9 Pro’s 128GB configuration by $700 to $850. Buyers who know they need 70B-plus dense models or large MoE models like GPT-OSS 120B, and who will actually use dual 10-gigabit networking to serve that model to other machines, should look at the Beelink GTR9 Pro instead, accepting its price premium as the cost of meaningfully better I/O. The Minisforum MS-S1 Max is the pick for buyers already inside the Minisforum ecosystem, or who specifically want a workstation-style chassis with room for more drives than the EVO-X2 offers, at a price that sits below the GTR9 Pro without requiring the EVO-X2’s mid-cycle tier hunting.
The bigger-picture point that applies regardless of which of the three you choose: this entire category exists because unified memory made large local models affordable relative to a discrete-GPU workstation, and the 2026 DRAM crisis has eaten a meaningful chunk of that affordability advantage. None of these machines are bad hardware. All three are more expensive, right now, than the underlying silicon justifies, and that is a memory-market problem rather than a design problem with any of the three boxes.
Frequently asked questions
Are the GMKtec EVO-X2, Beelink GTR9 Pro, and Minisforum MS-S1 Max actually the same computer inside?
They share the identical AMD Ryzen AI Max+ 395 processor and Radeon 8060S integrated GPU, so raw compute performance is similar across all three. The differences come from memory tier availability, storage slots, networking ports, cooling design, and price.
Do I need the 128GB memory tier?
Only if you plan to run models above roughly 30B parameters, particularly dense 70B-class models or large mixture-of-experts models like GPT-OSS 120B. For 7B-30B models, the 64GB tier on the EVO-X2 or MS-S1 Max is more cost-effective given current pricing.
Why are prices so much higher than they were earlier in 2026?
DRAM and NAND flash prices climbed sharply through 2026, and these machines are unusually memory-heavy relative to their total cost, so they have absorbed a larger percentage increase than most PC hardware categories. Valve’s documented 44 to 46 percent Steam Deck OLED price increase in May 2026 is a useful public data point for the scale of the shift.
Can these mini PCs replace a discrete GPU for gaming?
Not for demanding modern titles at high settings. The Radeon 8060S is a capable integrated GPU, but its performance sits well below a mid-range discrete card, and its main advantage is unified memory capacity for AI workloads rather than gaming frame rates.
How does token generation speed compare to a cloud API like ChatGPT or Claude?
Noticeably slower for large models. Single-digit to low-double-digit tokens per second on 70B-plus models is typical on this hardware, versus far higher throughput from cloud infrastructure built around dedicated high-bandwidth GPU clusters. Smaller models in the 7B-30B range feel much closer to cloud-speed responses.
Is ROCm support good enough for serious AI work on these machines?
ROCm support for Strix Halo has matured significantly through 2026 and works well with llama.cpp and Ollama’s Vulkan and ROCm backends, but it remains less universally supported across AI software than Nvidia’s CUDA ecosystem, which is a real consideration for workflows tied to CUDA-only tools.
Should I wait for prices to come down before buying?
Industry DRAM pricing forecasts referenced in coverage of the broader memory crisis suggest relief may not arrive meaningfully until 2028, so buyers who need the capability now should not assume a short-term price drop is coming, though it is worth comparing current listings across all three brands before committing given how much prices have moved week to week.
Which one is best for a first-time buyer entering local AI?
The GMKtec EVO-X2’s 64GB configuration offers the lowest financial commitment while still supporting a meaningful range of models, and its larger body of independent review data makes it the easiest of the three to verify performance claims against before buying.