Minisforum MS-S1 MAX-P495 lets devs run larger models on one box

The Minisforum MS-S1 MAX-P495 is now open for pre-order, priced at $7,399, with shipping expected in mid-October 2026. The standout feature isn’t the CPU or GPU clock speeds so much as the memory pool: 192GB of unified memory, with about 160GB assignable to the GPU. This is paired with a 2TB SSD, all housed in a compact 3.5-liter workstation chassis.

The machine runs on AMD’s Ryzen AI Max+ PRO 495, delivering 16 Zen 5 cores and 32 threads with a boost clock up to 5.2 GHz. The built-in Radeon 8065S GPU keeps 40 RDNA 3.5 compute units and climbs to 3.0 GHz. The NVIDIA-style neural processing unit (NPU) is rated at 55 TOPS, a 10% improvement over the prior model. The whole package is designed to handle large AI models locally, without sending data to the cloud.

Minisforum positions this as a big step for developers who want to run large models on premises for privacy or latency reasons. The 192GB memory, a first for this line, is the key differentiator here. Minisforum says memory capacity matters more than the modest CPU/GPU clock bumps when it comes to fitting and running very large models in real time.

In practice, Minisforum’s own tests show what the extra RAM buys you: a 284B-parameter model can run on a single unit with DeepSeek V4 Flash at about 12 tokens per second in Q4 quantization. Pairs of units can run even bigger models, and four-system clusters load the 380GB Q4_0 build of DeepSeek-R1 671B. The numbers are model- and workload-specific, but the takeaway is clear: memory capacity unlocks the ability to hold larger “experts” in memory and keep latency reasonable.

Price positioning puts the MS-S1 MAX-P495 in a premium tier. The pre-order price of $7,399 is 20% off a list price of $9,249 that Minisforum hasn’t charged yet. For users who don’t need that much memory, Minisforum also offers a 128GB configuration that remains a strong value for most AI-on-prem needs.