GMKtec has announced the upcoming release of its new flagship mini PC, the “EVO-X5 Pro,” scheduled for September 28, 2026, which is capable of running large language models with up to 300 billion parameters entirely locally. This development comes amid rising demand from enterprises and research institutions seeking to operate sophisticated autonomous AI agents in-house while mitigating security risks and bandwidth costs associated with transmitting data to external clouds.

The Impact of the EVO-X5 Pro Announced at IFA 2026

On September 5, 2026, at IFA 2026 in Berlin, Germany, Chinese mini PC manufacturer GMKtec officially unveiled its flagship compact AI workstation, the “EVO-X5 Pro,” announcing a release date of September 28, 2026. Powered by AMD’s latest high-end processor, the “Ryzen AI Max+ PRO 495,” this next-generation desktop AI platform supports an unprecedented unified memory capacity of up to 192GB. While traditional mini PCs were primarily designed for office tasks and light graphical processing, the EVO-X5 Pro integrates CPU, GPU, and NPU compute engines into a single, compact chassis. It enables inference for ultra-large language models to be completed entirely on a palm-sized desktop system without relying on data centers or cloud infrastructure. By executing advanced AI workflows completely offline—isolated from external networks—this device fundamentally redefines the boundaries of small-form-factor PCs and has captured global attention. Refer to the figure below.

The Significance of Running a 300B Parameter Model Locally

In recent years, the AI market has rapidly evolved toward “agentic PCs,” which autonomously plan and execute complex business workflows rather than simply generating text or answering basic queries. However, previous local environments (such as 128GB memory configurations) were limited to running 70-billion (70B) parameter models in 4-bit quantized versions or select 200B-class models. By expanding the memory capacity to 192GB, it is now possible to host Mixture-of-Experts (MoE) models with up to 300 billion (300B) parameters locally on a single system. This delivers significant advantages:
– Organizations can process sensitive customer data, medical records, and proprietary legal documents securely without transmitting them to external clouds.
– Network-induced latency is eliminated, delivering real-time responses.
– Ongoing operational costs associated with cloud AI monthly subscriptions and API usage fees are dramatically reduced.
For enterprises and government agencies seeking to balance confidentiality with cost containment, achieving full on-device AI capability holds immense strategic value.

The Technical Truth Behind the Unified Processor and 192GB Memory

Ryzen AI Max+ PRO 495 Architecture and Compute Performance

At the heart of the EVO-X5 Pro is AMD’s flagship APU, the “Ryzen AI Max+ PRO 495,” designed specifically for local AI inference and professional workloads. The processor features a unified architecture integrating three distinct processing engines on a single die:
– CPU Section: A 16-core, 32-thread configuration based on the Zen 5 architecture, with a maximum boost clock of 5.2GHz.
– GPU Section: AMD Radeon 8065S Graphics based on the RDNA 3.5 architecture, featuring 40 compute units and hardware ray tracing support.
– NPU Section: An AI accelerator based on the XDNA 2 architecture, delivering up to 55 TOPS of AI processing power on its own.
Complementing each other, these three engines achieve a combined maximum system AI performance of 131 TOPS. From standard data compute to complex code compilation, high-end 3D rendering, and real-time AI inference, the system offers exceptional processing power.

192GB Unified Memory Allocating 160GB to VRAM and Advanced Thermal Management

The defining technical breakthrough of this product is its support for up to 192GB of LPDDR5X unified memory running at 8533MT/s. Featuring a 256-bit memory bus, it achieves an ultra-wide memory bandwidth of approximately 273GB/s. Thanks to an architecture where the CPU and GPU share a single physical memory pool, up to approximately 160GB can be dynamically allocated as dedicated VRAM for graphics and AI processing. This enables the machine to load and run ultra-large AI models locally that were previously unexecutable on standard mini PCs or single discrete graphics cards due to VRAM limitations. To prevent thermal throttling under heavy workloads, GMKtec engineered a custom cooling system combining a large vapor chamber with three cooling fans. Users can select from three performance modes based on their workload, ensuring sustained processing performance even during prolonged AI inference sessions. Refer to the figure below.

Enterprise Reliability and the Future of Agentic PCs

Enterprise-Grade Design and Security Supporting 24/7 Continuous Operation

Designed for demanding business, medical, and governmental deployments, the EVO-X5 Pro is built for continuous, round-the-clock operation. Backed by rigorous quality validation from GMKtec, the machine guarantees durability for 24/7/365 uptime, while incorporating standard management and security features:
– AMD DASH remote management technology paired with a dedicated RJ45 DASH network adapter (enabling remote monitoring, OS updates, and BIOS configuration changes).
– A discrete physical security chip, dTPM 2.0 (protecting encryption keys and preventing unauthorized access).
– Comprehensive connectivity including dual USB4 ports on the front, and dual full-featured USB4 V2 ports, HDMI, dual LAN ports, and an OCuLink port on the rear.
By leveraging USB4 V2 or OCuLink, users can connect external graphics enclosures or link multiple EVO-X5 Pro units directly to build a high-bandwidth distributed computing cluster, offering excellent future scalability.

Cost Structure Revolution and Impact on the Local AI Market

To run massive 300B-class AI models locally, organizations previously needed custom desktop rigs equipped with multiple flagship discrete GPUs (such as NVIDIA RTX 5090s) or dedicated server infrastructure costing tens to hundreds of thousands of dollars. Beyond high upfront hardware costs, these setups presented major hurdles in terms of physical footprint and heavy monthly electricity bills. In contrast, the EVO-X5 Pro recreates an equivalent VRAM environment in a palm-sized form factor with remarkable energy efficiency, consuming only around 100W to 140W under load. Although targeted at professional price points, its Total Cost of Ownership (TCO) over time is extraordinarily economical. The launch of the EVO-X5 Pro represents a defining milestone that is set to accelerate the adoption of agentic PCs in late 2026 and beyond.

References

• [GMKtec Official Press Release] https://jp.gmktec.com/blogs/%E6%96%B0%E9%97%BB/gmktec-ifa-2026%E3%81%A7evo-x5-pro%E3%82%92%E7%99%BA%E8%A1%A8-%E6%9C%80%E5%A4%A73-000%E5%84%84%E3%83%91%E3%83%A9%E3%83%A1%E3%83%BC%E3%82%BF%E3%81%AEai%E3%82%92%E5%AE%8C%E5%85%A8%E3%83%AD%E3%83%BC%E3%82%AB%E3%83%AB%E3%81%A7%E5%AE%9F%E8%A1%8C%E3%81%97-agentic-pc%E6%99%82%E4%BB%A3%E3%81%B8
• [AMD Ryzen AI Max+ PRO 495 Official Product Specs] https://www.amd.com/en/products/processors/laptop/ryzen-pro/ai-max-pro-400-series/amd-ryzen-ai-max-plus-pro-495.html

【#MiniPC #GMKtec #LocalLLM #Ryzen #UnifiedMemory #AgenticPC #OnDeviceAI #AIWorkstation】