Editorial Team
5 min read
GMKtec's EVO-X5 Pro puts 192GB behind local AI—but not a clear speed claim
GMKtec's new EVO-X5 Pro mini PC pairs AMD's Ryzen AI Max+ PRO 495 with 192GB of unified memory, prioritizing the capacity to hold larger local models. The available evidence supports that capacity pitch, but not a comparable performance or price case yet.
- GMKtec's new flagship mini PC combines AMD's Ryzen AI Max+ PRO 495 with 192GB of unified memory; up to 160GB can be reserved for its integrated GPU.
- That capacity can hold a 300-billion-parameter model at four-bit quantization in theory, but it does not establish an interactive speed for a dense model.
- GMKtec has not announced a price, power-profile limits or reproducible benchmarks; reports also differ on whether its October availability was formally communicated.
GMKtec, a maker of compact PCs, has introduced the EVO-X5 Pro at IFA 2026 as a local-AI flagship. Its defining specification is not a discrete accelerator but 192GB of unified LPDDR5X memory paired with AMD's Ryzen AI Max+ PRO 495: an on-site account says up to 160GB can be assigned to the chip's 40-compute-unit Radeon 8065S integrated GPU.
For developers and researchers running models locally—the audience GMKtec is targeting—that allocation changes the largest model that can stay in fast memory. It does not settle the more practical question of how quickly the system can answer, how much it will cost, or which workload it is suited to.
Frandroid reports 192 Go of unified memory and up to 160 Go assignable to the GPU; its analysis estimates about 160 Go for a 300-billion-parameter model at four-bit quantization. Source: Frandroid.
More capacity, only a small bandwidth step
The Ryzen AI Max+ PRO 495 is a 16-core processor with integrated graphics, rather than a separate high-memory GPU. The analysis describes the processor as the same Zen 5 and Radeon 8065S design as the preceding 128GB configuration, with memory increasing from 128GB to 192GB and its data rate from 8,000 to 8,533 MT/s. That is a 50% increase in capacity but a 6.6% rise in stated memory rate; the article puts total bandwidth at 273GB/s, versus 256GB/s before.
That distinction is the limit on the marketing shorthand that the machine can run a 300-billion-parameter language model. The same analysis estimates that a model of that size, quantized to four bits, needs about 160GB—matching the claimed GPU allocation. It also estimates roughly 1.5 generated tokens a second for a model of that size because each output token requires reading its weights through that 273GB/s memory system. Those are analytical estimates, not a GMKtec benchmark, and they should not be read as a result for a sparse mixture-of-experts model, which activates only part of its parameters for each token.
