Minisforum has unveiled the MS-S1 MAX-P495, an upgraded version of its MS-S1 Max station, launched at the end of 2025. This compact machine weighs 2.8 kg and features an aluminum chassis, advanced six heat pipe cooling, and two 10 GbE network ports. It is powered by the Ryzen AI Max+ PRO 495 processor and includes 192 GB of LPDDR5X-8533 memory, with 160 GB allocated to the GPU. The device is designed to fit into a 2U rack and can be stacked to form a cluster, enabling improved performance when multiple units are used together. The MS-S1 MAX-P495 is expected to be available for purchase before the end of the month, though its price remains undisclosed. The earlier version with 128 GB of memory was sold for around 2,500 euros, and it is unlikely that the newer version with more memory will be cheaper. The product specifications highlight the device's ability to run several large language models. For example, the DeepSeek V4 Flash model, which has 284 billion parameters, can process 15 tokens per second on a single machine. When two MS-S1 MAX-P495 units are connected, the Qwen3.5-397B model, with 397 billion parameters, can run at 16 tokens per second. Additionally, a cluster of four units can run the DeepSeek-R1 671B model, which is 380 GB in size. These performance figures are described as internal measurements and are yet to be independently verified. Minisforum claims that the MS-S1 MAX-P495 offers 131 TOPS (trillion operations per second) of AI computing power, combining the 55 TOPS from the NPU (neural processing unit) and the estimated contribution from the Radeon 8065S GPU. The company also announced the N5 MAX-P495, a five-bay network-attached storage (NAS) device that can hold up to 200 TB of data. It features the same processor and 192 GB of memory as the MS-S1 MAX-P495 and includes pre-installed software called OpenClaw, along with support for Hermes Agent, two open-source frameworks for autonomous agents that can run locally or remotely. The MS-S1 MAX-P495 is particularly notable for its performance with large language models. The ability to process 15 tokens per second on the DeepSeek V4 Flash model suggests a smooth and efficient reading speed. The 16 tokens per second for the Qwen3.5-397B model using two machines is more surprising, as clusters connected via 10 GbE networks often face synchronization challenges. Minisforum suggests that two units could match the performance of a high-end Mac Studio M5 Ultra with 512 GB of memory, but likely at a lower cost and with more space on the desk. A setup using four machines to run the DeepSeek-R1 671B model is described as a demonstration, with the company stating the model "ran successfully" but not specifying the speed. The estimated cost for this setup, including a 10 GbE switch, is around 12,000 euros. Minisforum has positioned the NAS as a suitable platform for autonomous agents, given the availability of storage and computing resources during off-hours.