Minisforum MS-S1 MAX-P495 Turns 160GB Into AI VRAM

A Mini Workstation Built Around One Number: 160GB

Minisforum used IFA 2026 in Berlin to show two machines built on AMD's Ryzen AI Max+ PRO 495, and the specification that matters for local AI is not the TOPS figure. It is the memory allocation. Both systems carry up to 192GB of LPDDR5X running at 8533 MT/s, and up to 160GB of that pool can be handed to the integrated GPU as graphics memory.

• AI Mini Workstation MS-S1 MAX-P495: Ryzen AI Max+ PRO 495, up to 192GB LPDDR5X-8533, up to 160GB allocatable as VRAM, dual M.2 up to 16TB
• AI Agent NAS N5 MAX-P495: the same compute paired with up to 200TB of local storage for vector databases, embeddings and persistent agent workloads
• AI compute: up to 131 TOPS total, including up to 55 TOPS from the XDNA 2 NPU
• On show: IFA Berlin Hall 6.2-104, September 4 to 8, 2026. Minisforum has not published retail pricing for either MAX-P495 model

What Is Inside the Ryzen AI Max+ PRO 495?

The chip is the "Gorgon Halo" part AMD introduced with the Ryzen AI Max PRO 400 series in May 2026, and it is a meaningful step up from the Ryzen AI Max+ 395 that has been powering the current wave of Strix Halo mini PCs.

You get 16 Zen 5 cores and 32 threads boosting to 5.2GHz with 80MB of total cache, and a configurable TDP spanning 45W to 120W. Graphics come from a Radeon 8065S with 40 compute units. The memory controller is the part that makes this a local-AI machine rather than a fast desktop: a 256-bit interface feeding LPDDR5X-8533, with a 192GB ceiling.

That combination is why the 160GB VRAM allocation is possible at all. On a conventional desktop, graphics memory is whatever the discrete card shipped with — 24GB is a lot, 32GB is unusual. On a unified-memory design the split is a firmware setting, so a 192GB machine can present more addressable model memory than any consumer graphics card on the market.

What Can You Actually Run on 160GB?

This is the practical question, and the answer is most of the open-weight catalogue at sensible quantisation levels. A 27B-class model at 4-bit lands around 16 to 18GB. A 120B-class mixture-of-experts model at 4-bit fits comfortably. The 3-bit quantisation of GLM-5.3-Flash that we tested against a 128GB budget last month has 32GB of headroom to spare here, and there is room left for a long context window on top.

What 160GB does not get you is the very top of the open-weight range. Trillion-parameter mixture-of-experts models still want 240GB and up even at aggressive quantisation. That is a real ceiling, but it sits well above where most self-hosted work actually happens.

The MS-S1 MAX-P495 backs the compute with dual M.2 slots supporting up to 8TB each in RAID 0 or 1, a full-length PCIe x16 slot wired for PCIe 4.0 x4, dual 10GbE, USB4 V2 and HDMI 2.1 FRL. Cooling is six heat pipes, dual turbine fans and a phase-change thermal interface, with a 320W supply feeding a 120W sustained and 160W peak processor envelope, and four operating profiles including a rack mode.

Why Pair 131 TOPS With 200TB of Storage?

The N5 is the more interesting of the two, because it answers a question the mini PC market has mostly ignored. Running a model locally is only half of a private AI setup. The other half is the corpus — the documents, embeddings and vector indexes that a retrieval-augmented pipeline searches before the model ever sees a prompt.

Minisforum's pitch for the N5 MAX-P495 is that both halves live in the same box: up to 200TB of local storage sitting directly next to the accelerator that will query it, with no network hop and no cloud round trip. For a small team indexing its own archive, that is a genuinely different architecture from bolting a GPU onto a general-purpose NAS.

When Can You Buy One?

Not yet, and that is the caveat worth stating plainly. Minisforum showed both machines at IFA with live demonstrations and store links, but has not published retail pricing or firm ship dates for the MAX-P495 configurations. The existing MS-S1 MAX with the older Ryzen AI Max+ 395 and 128GB of memory lists at $2,919.90, which gives you a floor rather than a forecast.

The wider point is that IFA 2026 turned into a local-AI hardware show. GMKtec is debuting its own Ryzen AI Max+ PRO 495 machine, the EVO-X5 Pro, at the same event, and Acer brought the Veriton RI110 we covered this week. Three vendors, one week, all building desk-sized boxes whose main selling point is how much model you can fit in memory. More on the hardware in our mini computer coverage.

Sources: MINISFORUM press release — Next-Gen Edge AI Computing Solutions at IFA 2026 — September 2026; AndroidGuys — September 2026; Notebookcheck — Ryzen AI Max+ PRO 495 specifications — 2026.