Minisforum has revealed the MS-S1 MAX-P495, a compact workstation-class mini PC aimed at local AI and built around the AMD Ryzen AI Max+ PRO 495 processor. The company presents it as a distinct new model rather than a firmware update to the MS-S1 MAX. Its announcement post is dated 12 November 2026.
Minisforum calls the machine 'upcoming' and says it is 'almost here', but gives no release date. It also says initial production runs will be 'highly limited' and invites readers to join an early-access newsletter for pricing, benchmarks and an early-bird discount.
What the specification says
The memory is listed as LPDDR5x, and Minisforum's how-to guide specifies LPDDR5X-8533. The guide describes setting the VRAM allocation in BIOS through the UMA Specified and UMA Frame Buffer Size options. Minisforum specifies approximately 273 GB/s of memory bandwidth, 120 W sustained-load power and 160 W peak power.
| Memory | 192 GB LPDDR5X-8533 (MS-S1 MAX: 128 GB) |
|---|---|
| Max VRAM allocation | 160 GB (MS-S1 MAX: 96 GB) |
| NPU rating (Minisforum) | 55 TOPS (MS-S1 MAX: 50 TOPS) |
| System AI compute (Minisforum) | 131 TOPS (MS-S1 MAX: 126 TOPS) |
| Networking | Dual 10GbE (Realtek RTL8127), Wi-Fi 7, Bluetooth 5.4 |
| Ports | HDMI 2.1 (8K@60), USB-A, two USB4 (40 Gbps), two USB4 V2 (80 Gbps) |
| Expansion | Internal PCIe x16 slot, PCIe 4.0 x4 electrical |
The comparison figures against the standard MS-S1 MAX come from Minisforum's own table, as do the TOPS ratings. The 131 TOPS number combines the CPU, integrated Radeon graphics and NPU, so it is a system-level rating rather than an NPU figure. The internal slot is meant for SSDs, network adapters or controller cards, and Minisforum says it does not support discrete graphics cards.
Local model claims
Minisforum says the P495 can run suitable quantised 70B and 120B-class models locally. It adds that real performance depends on the model, quantisation, context length and software, and that the KV cache and runtime overhead need memory on top of the model weights.
The company also says it tested DeepSeek V4 Flash 284B at Q4 quantisation, using about 144 GB of model memory, and measured 12.35 tok/s generation. That is a vendor-run result, and no methodology, inference engine or context length is given. Minisforum's guide also says the 55 TOPS NPU is not used automatically by ordinary Ollama or llama.cpp inference; it needs compatible software and supported operations.
Why it matters
This is our reading of the vendor's figures: a 160 GB VRAM ceiling, against 96 GB on the standard model, could let the P495 hold larger quantised models. Token speed is likely to stay bound by the roughly 273 GB/s of memory bandwidth, so more capacity does not necessarily mean faster generation. All the material so far comes from Minisforum's own blog, with no independent benchmarks and no AMD confirmation of the processor's specifications.
What we do not know yet
- What the MS-S1 MAX-P495 will cost and when it will go on sale
- Whether AMD's own specifications confirm the Ryzen AI Max+ PRO 495 memory capacity, NPU rating and 131 TOPS figure
- Which inference engine, context length and settings produced the 12.35 tok/s DeepSeek V4 Flash result, and whether independent testers can reproduce it
- How thermals and noise compare with the standard MS-S1 MAX under sustained load
- How performance compares with the standard MS-S1 MAX and with multi-GPU desktops in independent tests





