Minisforum P495 hardware unveiled at IFA 2026 brings AMD’s Ryzen AI Max+ PRO 495 to two compact systems: the MS-S1 MAX-P495 workstation and N5 MAX-P495 network-attached storage device. Their headline feature is support for up to 192GB of unified memory, giving local AI workloads substantially more working space than ordinary mini PCs.
- Both systems pair a 16-core Ryzen AI Max+ PRO 495 with up to 192GB unified memory.
- The workstation targets local inference and expansion; the NAS combines AI compute with large local storage.
- Model-size claims need configuration, software and measured-performance context.
What the Minisforum P495 launch includes
Minisforum announced both products on September 4. The MS-S1 MAX-P495 is a compact workstation with Radeon 8065S graphics, a dedicated NPU and a full-length PCIe expansion slot. The N5 MAX-P495 is positioned as an “AI Agent NAS” that can combine local model execution with up to 200TB of storage, according to the company.
The practical significance of the Minisforum P495 systems is memory capacity: a unified pool can hold models that do not fit in the dedicated graphics memory of mainstream PCs, although usable speed depends on quantisation, software support and workload. A large capacity figure is not the same as a guarantee that every advertised model will run at an acceptable response rate.
Minisforum P495 facts table
| Specification | Launch claim |
|---|---|
| Processor | AMD Ryzen AI Max+ PRO 495, 16 cores and 32 threads |
| Total AI compute | Up to 131 TOPS across processing engines |
| Memory | Up to 192GB unified memory; up to 160GB allocatable as graphics memory |
| MS-S1 networking | Dual 10Gb Ethernet, Wi-Fi 7 |
| Availability | Final pricing and shipment details were not complete at announcement |
Tom’s Hardware reported that the N5 configuration can store local agent data and run software including OpenClaw or Hermes Agent, while the MS-S1 is aimed more directly at desktop AI computing. TechRadar highlighted the workstation’s cooling system, rack-ready form and connectivity. These are useful implementation details, but neither outlet published full benchmarks in its launch coverage.
Why unified memory changes local inference
Large language models consume memory for weights and working context. A unified architecture lets CPU and GPU functions address the same pool, reducing the hard separation between system RAM and discrete graphics memory. That can make unusually large quantised models feasible on a compact system.
The trade-off is that capacity, bandwidth and tokens per second remain different measures. Minisforum says a two-system cluster can run very large models locally, but buyers should wait for reproducible tests that state model build, quantisation, context length, power mode and output speed. The distinction echoes our coverage of Nvidia’s IFA local-AI routing, where orchestration matters as much as silicon.
Local processing can also reduce how much sensitive material must leave an organisation, but it does not create governance automatically. Teams still need authentication, model provenance, audit logging and patching. Our analysis of enterprise AI control planes explains why ownership of compute does not remove operational risk.
Who should consider these systems
The MS-S1 MAX-P495 fits developers, small research teams and creators who need an unusually large shared memory pool without installing a rack server. The N5 MAX-P495 is more relevant where source data already lives on network storage and local inference can be placed beside it.
Prospective buyers should compare measured sustained performance, noise, power draw, warranty coverage and memory configuration. Minisforum’s launch establishes a compelling specification class; independent testing must establish value.
FAQs
What are the new Minisforum P495 systems?
They are an MS-S1 compact workstation and N5 AI-focused NAS built around AMD’s Ryzen AI Max+ PRO 495 processor.
Can they run large AI models locally?
Their memory capacity is designed for that purpose, but practical model size and speed depend on configuration, quantisation, software and context length.
How much do they cost?
Complete launch pricing and shipping details were not established in the sources reviewed for this package.
A useful purchasing test is to match the intended model size, quantisation and context length to published memory requirements before choosing a chassis. Buyers should also separate peak hardware specifications from sustained application performance. Cooling, software support and storage throughput can determine whether the Minisforum P495 platform delivers its practical advantage during long local-AI sessions.
Sources
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
