Beelink GTR9 Pro (Ryzen AI Max+ 395) review

If you’re stuck between the 128GB Strix Halo boxes, the chip won’t settle it. They all run the same Ryzen AI Max+ 395, so what you’re paying for is the box. The Beelink GTR9 Pro wraps that chip and 128GB of soldered memory in a case with dual 10GbE, a 230W PSU built in and a full vapour chamber. At the time of writing that’s $4,199, about $700 more than the GMKtec EVO-X2, and on some units those 10GbE ports have dropped out.

The take

Anyone who priced up a GTR9 Pro in October 2025 found it often selling for $1,999 or less, roughly what the GMKtec EVO-X2 cost. At the time of writing Beelink sells one configuration, 128GB with a 2TB SSD, at $4,199 on pre-order, and Amazon US has it in stock at $4,299. The 128GB GMKtec is $3,499.99, which puts the Beelink about $699 above it, or about $549 against the GMKtec’s own 2TB option. The Framework Desktop is $3,449 as a DIY kit, and out of stock. The memory is soldered LPDDR5X, which the chip needs for its bandwidth, and it can never be upgraded. Capacity is settled at checkout, and it’s the same 128GB the GMKtec has. So the extra money goes on the ports, the cooling and the PSU inside, and it’s worth paying only if you’ll plug into the 10GbE or want the box quiet.

Dual 10GbE is the reason to pick it over the GMKtec and the Framework, which have a single 2.5GbE and a single 5GbE port. The 10GbE earns its place with a model library kept on a NAS and pulled across the network. Pulling gpt-oss-120b (about 61GiB) at line rate, before any overheads, takes about 52 seconds over 10GbE, against 1 minute 45 over 5GbE and 3 minutes 30 over 2.5GbE. At its rated 7,000 MB/s, the fitted 2TB SSD reads it in about 9 seconds, so the network is for models that don’t live on the box.

That needs the ports to stay up, and on some units they haven’t. Beelink’s support blamed the Ethernet driver in September 2025. ServeTheHome, whose unit had Intel’s E610 controller, saw the ports work at first and then vanish after updates, suspected firmware and was told a fix was coming. The forum thread on the ports failing under load was still active on 12 November 2026. On Linux the same ports need an extra step, since owner Roberto Brunialti found their drivers aren’t in the kernel by default, though you can download them.

At 180 x 180 x 90.8mm and about 2.3kg it’s big for a mini-PC, with dual fans and a full vapour chamber inside. ServeTheHome says most of that chassis is cooling, which is what lets the box run at Beelink’s 140W rating at low noise. The reviewers there wouldn’t call the Beelink silent at around 120W, but they measured 39-41 dBA at about 125-128W and still found it the quietest of their four Strix Halo boxes. Tame That Tech reports noise as low as 38 dB at 140W. Under load the box draws up to about 180W at the wall, which is inside the 230W rating of the PSU built into the all-metal case. The GMKtec runs the same chip from an external charger of about the same rating, so the Beelink leaves no power brick on the desk. It’s a premium build, and Beelink backs it with a 3-year warranty. The GTR9 Pro suits a desk where fan noise would bother you. If you keep Windows, bear in mind that ServeTheHome’s review unit came with EdrawMind, EdrawMax and Filmora preinstalled, and they wouldn’t come out through Add or Remove Programs. The reviewer came away trusting the Windows image less.

What it’ll run

Whether a model fits comes down to how much of the 128GB the graphics core gets, since it’s unified memory (one pool shared by the processor and graphics core). The box ships with 96GB on the graphics side, and set that way the graphics core can reach 112GB. That 96GB is as much as an RTX Pro 6000 carries, though the card has about seven times the memory bandwidth. Bandwidth (how fast the chip can read the model out of memory) is what sets the pace when a model generates tokens, and the 395 is rated at 256 GB/s with about 215 measured. So a dense model has a rough ceiling of bandwidth divided by file size, and for a 70B at Q4_K_M, a 42.52GB file, that’s about 5 tok/s. ServeTheHome measured 4.89 on Llama 3.3 70B, without stating the quant. gpt-oss-120b ran at 31.41 tok/s because it’s a mixture-of-experts (MoE) model, where only 5.1B of its 117B parameters are active per token. Beelink’s own figure for Qwen 32B at Q8 is 6.23 tok/s, about the pace of the 70B. For anything you’ll sit and wait on, load an MoE model like gpt-oss-120b, and keep a dense 70B for jobs that can take about six times as long.

There’s no CUDA here, and that’s true of every Strix Halo box, not just this one. Pretty much everything in AI works on Nvidia out of the box, though ServeTheHome expects AMD’s work on ROCm 7 to narrow that gap. The GTR9 Pro ships with Windows 11 Pro, and Beelink quotes 50 TOPS for the NPU and 126 TOPS for the chip as a whole. Of the owners who’ve run one, Artur Wróblewski reports that Ollama works fine on Ubuntu 24.04, and Dejan, without saying which OS, recommends the Lemonade AI server as the backend.

Who should buy it

It makes sense if your models already live on a NAS with 10GbE, since that’s the job the ports are for. The other job is a two-box experiment in clustering (several boxes working on one model over the network). AMD documents it over a 10Gbps link on other 395 machines, though on this box it’s Beelink’s claim and no reviewer has tested it. Both jobs stop if those ports drop out. A buyer who only wants it ready-built and quiet, with the PSU inside and a 3-year warranty, isn’t relying on them at all.

If your work depends on CUDA, skip every Strix Halo box, because a GB10 machine like the DGX Spark gets you 128GB of unified memory on Nvidia. With nothing to plug into the 10GbE, the same chip costs less at the time of writing in the GMKtec EVO-X2 at $3,499.99, or the Framework Desktop at $3,449 as a DIY kit once it’s back in stock. And if you need those ports dependable from day one, hold off for now, because the forum thread on the fault was still getting posts on 12 November 2026.