GMKtec EVO-X2 Review: 128GB Strix Halo Now $3,499.99, Up From $1,999

GMKtec's US store listed the 128 GB EVO-X2 mini PC at $3,499.99 with a 1 TB drive when read on Nov. 11, 2026, up from the $1,999 that Notebookcheck reported for a 128 GB model in September 2025, a listing that did not state the drive size. GMKtec's spec page lists the same Ryzen AI Max+ 395 and LPDDR5X-8000 memory the 2025 reviews describe, the same chip as the Framework Desktop in our Nov. 4 review, in a 193 x 185.8 x 77 mm box with an external 230 W adapter.

We found three independent reviews as of Nov. 11, 2026. ServeTheHome published its review of a 128 GB, 2 TB unit on July 11, 2025, with LLM speed, wall power and noise. Notebookcheck published a 64 GB, 1 TB review on Sept. 2, 2025, with power and noise measured at 15 cm. Tom's Hardware published a 64 GB, 1 TB review on Feb. 4, 2026, with CPU, gaming and thermal results. Of the three, only ServeTheHome reported LLM speed, and its review does not name the inference tool. That limits what any buyer can conclude about this box against rivals measured on gpt-oss-120b.

What's New: EVO-X2 Specs, Memory and Price

The EVO-X2 is a mini PC built on AMD's Ryzen AI Max+ 395, which GMKtec lists with 16 cores and 32 threads, a 40-core RDNA 3.5 Radeon 8060S GPU and an XDNA 2 NPU rated at 50 TOPS. ServeTheHome called it the first Ryzen AI Max+ 395 system it had reviewed. GMKtec rates the chip at 120 W sustained and 140 W peak in this chassis.

The memory is the point. GMKtec sells 64 GB or 128 GB of LPDDR5X-8000, soldered and not upgradeable. The CPU and GPU share it, so the 128 GB model can expose up to 96 GB to the GPU, according to Tom's Hardware; ServeTheHome ran both a 64/64 and a 32/96 split. GMKtec does not publish a memory bandwidth figure. AMD lists 256 GB/s for the Ryzen AI Max+ 395 in its Ryzen AI Halo developer platform, and Framework gives the same figure for its desktop.

The price is the newest fact. On Nov. 11, GMKtec's US store listed the 128 GB model at $3,499.99 with 1 TB and $3,649.99 with 2 TB, each against a $3,999.99 crossed-out price, and the 64 GB, 1 TB model at $2,199.99. Every variant was listed as available, under a banner warning of a coming price increase. In February 2026, Tom's Hardware reported the 128 GB, 2 TB configuration at $2,199 from GMKtec and $2,299 at Micro Center; Tom's itself reviewed the 64 GB unit. Like for like on the 2 TB model, the store price is about 66% higher than in February, by our arithmetic.

What It Solves: A 70B Model on the EVO-X2's GPU

The 128 GB EVO-X2 lets a dense 70B model run entirely on the integrated GPU. ServeTheHome ran Llama 3.3 70B Instruct at Q6_K, a load that needed more than 50 GB of GPU memory, at 3.7 to 3.8 tokens per second fully on the Radeon 8060S. Running on the CPU, it saw only around 2 tokens per second. Qwen3-32B ran at 10.4 to 10.5 tokens per second; ServeTheHome's text does not state that model's quantization. Each run answered one short question, on the software of mid-2025.

Memory bandwidth sets that pace. Generating text from a dense model means reading its weights from memory for every token, so speed depends on how fast memory delivers them, not only on compute. Mixture-of-experts models such as gpt-oss-120b read only part of their weights per token, which helps explain why ServeTheHome measured about 45 tokens per second on gpt-oss-120b (MXFP4) in LM Studio on the Framework Desktop. That is a different model, quantization, tool and machine from the Llama 3.3 70B Q6_K run, so the two speeds show the trade-off, not a ranking of the boxes. We found no published gpt-oss-120b measurement of the EVO-X2 as of Nov. 11, 2026.

GMKtec makes its own claims. Its product page, read on Nov. 11, 2026, says the 128 GB model runs Qwen3 235B at an average of 11 tokens per second and lists gpt-oss-120b at 19.25 tokens per second, without naming the quantization, the context length or the inference engine. Those are vendor figures, we found no independent measurement of them, and they cannot be set against the LM Studio figures other outlets report for gpt-oss-120b.

EVO-X2 Pros and Cons

The strongest pro is memory per box. ServeTheHome, which reviewed the 128 GB unit, said one test prompt got a good answer from the 70B model and nonsense from many smaller models, and that needing more than 50 GB of GPU memory takes such a model out of the class of GPUs like the NVIDIA L40S. On its 64 GB unit, Tom's Hardware found the CPU results almost indistinguishable from the Framework Desktop in Handbrake and Geekbench, and Notebookcheck found performance on a par with the Bosgame M5, another Ryzen AI Max+ 395 mini PC. On a 64 GB unit, Notebookcheck measured 0.65 W off, 6.5 to 10.5 W idle, 84.3 W average load and a 214.9 W peak at the wall.

The cons are noise in two of the three reviews, networking and price; Notebookcheck itself called the machine comparatively cool and quiet under full load. ServeTheHome measured 37 to 38 dBA at idle and 41 to 43 dBA during Llama 3.3 70B inference in a studio with a 34 dBA noise floor, and 170 to 180 W while gaming. Notebookcheck, with a microphone 15 cm away, measured 27.4 to 27.5 dB(A) idle and 47.5 dB(A) under load; the two methods differ, so the figures should not be set side by side. Tom's Hardware heard the fans across a living room during CPU benchmarks. ServeTheHome wished for 10GbE at this price, and that was at $1,999.

Who the EVO-X2 Is For

It fits a developer who wants a dense 70B model at Q6 on the desk, at a few tokens per second, without a discrete GPU. It fits a team that needs the box soon: GMKtec listed every configuration as available on Nov. 11, while Framework's US store showed every Max+ 395 configuration of its desktop out of stock on Nov. 4; check stock before ordering. It also fits a buyer who wants CPU work and 1080p gaming from the same machine; on its 64 GB unit, Tom's Hardware measured 57 FPS in Metro Exodus at 1080p with RTX settings.

It does not fit a buyer chasing tokens per second for the money. ServeTheHome called it a really strong system for AI but said $2,000 felt like a lot for basic desktop use or gaming, and Tom's Hardware called the 64 GB unit pricey at $1,499; neither has revisited that judgment at $3,499.99, and a buyer who needs measured gpt-oss-120b speed has better-documented options. It also does not fit anyone who needs to cluster boxes: its only Ethernet port is 2.5GbE.

How the EVO-X2 Compares With Framework Desktop and Ryzen AI Halo

Against the Framework Desktop, the hardware is a wash. Same chip, same 128 GB of LPDDR5X-8000, and Tom's Hardware found CPU results almost identical. Framework listed its 128 GB DIY Edition at $3,449 on Nov. 3, out of stock, as a system only, and a 2 TB SSD adds $505; GMKtec's 128 GB model with a 1 TB drive lists at $3,499.99, so the two prices are not like for like. In its separate Framework review, ServeTheHome measured 39 to 41 dBA loaded, against 41 to 43 dBA on the EVO-X2 during 70B inference; the rooms and loads differ. Only the Framework has a gpt-oss-120b figure; our Framework review has the details.

The HP Z2 Mini G1a uses the Ryzen AI Max+ PRO 395 with an internal 300 W supply, and Notebookcheck cites a power limit of up to 160 W for it. Notebookcheck rated the EVO-X2 slightly better overall than the HP despite that higher limit. On gpt-oss-120b in LM Studio, StorageReview measured close to 40 tokens per second on the HP and HotHardware about 24, as our HP Z2 Mini G1a review explains. Those figures are for a different model from ServeTheHome's Llama 3.3 70B Q6_K run on the EVO-X2, and ServeTheHome did not name its tool, so the numbers cannot be compared directly.

AMD's Ryzen AI Halo developer platform, built on the Ryzen AI Max+ 395 with 128 GB in the unit StorageReview tested, was priced at $3,999 with 2 TB in July 2026. StorageReview measured 222 tokens per second of total throughput on gpt-oss-120b in vLLM at batch 64, with 256 input and 256 output tokens, against 701 for the DGX Spark. That is serving throughput across many requests, not single-user speed like the figures above; our Ryzen AI Halo review covers that result. NVIDIA's DGX Spark has 128 GB at 273 GB/s and listed at $6,950 with 4 TB on NVIDIA's marketplace on Oct. 31; our DGX Spark review sets out where its speed lead holds. The EVO-X2 has no batch or vLLM measurement, so the table under this story shows it against these rivals on memory and price, not on speed.

What to Watch Before Buying

The first date to watch is GMKtec's own: its store warns of a price increase but gives no date, so the $3,499.99 price may not hold through November 2026. The second is a gpt-oss-120b measurement of the EVO-X2 from a recognized lab, with the model, quantization and tool named, which would let buyers set it against the Framework Desktop and the Ryzen AI Halo. The third is Framework's Max+ 395 restock, for which Framework gave no date on Nov. 4. ServeTheHome wrote that AMD said ROCm 7 and later releases would support the chip's RDNA 3.5 GPU, after its own testing was done, so newer software may move the 2025 figures; none of the reviews we read has retested this machine since.