MINISFORUM, a global edge computing solution brand and system integrator, today unveiled the AI Mini Workstation MS-S1 MAX-P495 at IFA 2026, alongside the AI Agent NAS N5 MAX-P495. Powered by the AMD Ryzen™ AI MAX+ PRO 495 processor, the MS-S1 MAX-P495 represents a significant leap in local AI computing, delivering up to 131 TOPS of total AI performance, 192GB of unified memory, and support for up to 160GB of graphics memory. The company confirmed that the workstation will go on sale within September 2026.
Flagship AI Performance
At the heart of the MS-S1 MAX-P495 lies the AMD Ryzen™ AI MAX+ PRO 495 – the flagship model of the Ryzen AI Max PRO 400 series. The processor features 16 cores and 32 threads based on the Zen 5 architecture, with clock speeds reaching up to 5.2GHz. It integrates Radeon™ 8065S graphics and a next-generation NPU delivering up to 55 TOPS of dedicated acceleration. The total system AI compute reaches 131 TOPS, up from 126 TOPS in the previous generation.
The workstation supports up to 192GB of LPDDR5x-8533 unified memory, with up to 160GB allocatable as VRAM. This represents a substantial upgrade from the 128GB memory capacity of the earlier MS-S1 MAX-395 model. The unified memory architecture delivers 273GB/s peak bandwidth, enabling seamless sharing of the memory pool between CPU and GPU.
Running 284B-Parameter Models Locally
The MS-S1 MAX-P495 sets a new benchmark for single-system AI capabilities. In real-world local AI deployments, the workstation can run the 284B-parameter DeepSeek V4 Flash model with Q4 quantization at an inference speed of approximately 15 tokens per second. For 122B-class models with Q8 quantization, it achieves speeds of approximately 27 tokens per second. The system also supports Q8 high-precision quantization with 131K full-context retention.
Compared to the previous-generation MS-S1 MAX-395, which could load 122B models with Q8 quantization, the P495 variant increases single-system model capacity from 122B to 284B.
Designed for Real-Time AI Workloads
Engineered for intensive local AI tasks in an ultra-compact footprint, the MS-S1 MAX-P495 is purpose-built for low-latency AI applications including AI agents, computer vision, generative AI, and other real-time inference scenarios. It delivers responsive local inference with millisecond-level latency, making it suitable for demanding model inference, AI development, and computationally intensive workloads.
The system features dual M.2 NVMe slots supporting up to 16TB of total storage, dual 10GbE wired networking, and Wi-Fi 7 connectivity. Front and rear USB4 and USB4 V2 interfaces, along with a PCIe x16 expansion slot, provide flexible connectivity for peripheral devices.
Premium Thermal Design
The MS-S1 MAX-P495 employs an aerospace-grade aluminum chassis with a diamond-matrix grill for robust durability and efficient airflow. The cooling system combines a pure copper heat spreader with a six heat-pipe array and dual turbine fans, utilizing advanced phase-change material for superior thermal transfer. The system supports 120W sustained power delivery with 160W peak PPT, ensuring stable performance under long-term high-intensity computing workloads. A built-in 320W high-efficiency power supply with CCC, FCC, UL, CE, UKCA, and other global certifications ensures reliable power delivery.
A Complete Edge AI Ecosystem
The MS-S1 MAX-P495 and N5 MAX-P495 form a complete edge AI computing system – the MS-S1 serves as the high-performance AI compute frontend while the N5 MAX-P495 functions as the backend integrating compute, massive storage (up to 200TB), data, and AI knowledge. This integrated architecture provides a powerful foundation for real-time inference, persistent AI workloads, and scalable local AI applications.
With the MS-S1 MAX-P495, MINISFORUM continues its commitment to democratizing AI computing by bringing powerful local AI capabilities to a broader range of users and real-world applications. The workstation addresses the growing demand for local, private AI deployment across government and enterprise sectors, supporting offline privatization, heavy model fine-tuning, multi-user concurrency, and industrial 24/7 edge AI scenarios.
