Unified memory decides which models a mini PC can run. Buy the tier you need before Black Friday prices move.
Prices change often during Black Friday, so check the live price on each product page.
The one pick that clears our deal bar today (active, USD, price-checked within 24 hours, no inflated strikethrough) is the Beelink SER9 PRO+, a Ryzen 7 H 255 box with 32 GB of LPDDR5X for 3-8B models. If speed on 7-14B models matters more than the form factor, a desktop with a 12 GB GPU still wins; see the DIY alternative.
By Mike Perry · Updated November 12, 2026
Black Friday is usually the week mini PC brands cut prices across their lineups, and Strix Halo boxes are the main reason local-LLM buyers care. A mini PC with 128 GB of unified memory can load models that no single consumer graphics card can hold, at a fraction of a multi-GPU workstation's power draw. This year, listings for those boxes have climbed instead of falling, so the right move for most 128 GB buyers is to set a target price and wait.
Wait if you're unsure whether you need 64 GB or 128 GB, because memory on every machine below is soldered and can't be upgraded later. Also wait if you mostly run 7-14B models and want maximum speed: a desktop with a discrete GPU generates those faster for less money.
The one spec that decides everything is unified memory. On these APUs, the integrated GPU borrows system RAM as VRAM, so the memory you buy is the largest model you can ever run. Bandwidth then sets how fast it runs. This guide anchors on spec tiers and launch prices, not deal prices. The live price button on each product page shows the current figure.
| Unified memory | Practical model ceiling (Q4) | Example models | Picks in this guide |
|---|---|---|---|
| 32 GB | 7-14B, with room for the OS | Llama 3.1 8B, Qwen3 14B | Beelink SER9 PRO+ (deal pick) |
| 64 GB | 27-32B | Gemma 3 27B, Qwen3 32B | Wait list: MINISFORUM N5 MAX |
| 128 GB | 70B-class, gpt-oss 120B | Llama 3.3 70B, gpt-oss 120B | Wait list: Beelink GTR9 Pro, MINISFORUM MS-S1 MAX, GMKtec EVO-X2 |
A useful rule of thumb: a Q4 model needs roughly 0.6 GB per billion parameters, plus a few GB for context and the operating system. Mixture-of-experts models like gpt-oss 120B are the exception that makes 128 GB boxes appealing. They need the memory to hold every expert, but they read only the active ones per token, so they generate much faster than a dense model of the same size.
| Machine | Best For | Key Spec | Status on Nov 12 | Verdict |
|---|---|---|---|---|
| Beelink SER9 PRO+ | 3-8B always-on assistant | Ryzen 7 H 255, 32 GB LPDDR5X, USB4 | Deal-ready | Best value pick today |
| Beelink GTR9 Pro | 70B-class and gpt-oss 120B | Ryzen AI Max+ 395, 128 GB, dual 10GbE | Far above launch pricing | Wait |
| MINISFORUM MS-S1 MAX | 128 GB plus a PCIe slot | Ryzen AI Max+ 395, 128 GB, 320 W PSU | Far above launch pricing | Wait |
| MINISFORUM N5 MAX | NAS plus 27-32B inference | Ryzen AI Max+ 395, 64 GB, 5 drive bays | Above launch pricing | Wait |
| GMKtec K8 Plus | 3-8B models | Ryzen 7 8845HS, 32 GB DDR5 | Listing up since September | Wait |
Verdict: The 32 GB mini PC to buy this week for 3-8B models, home-assistant and small RAG workloads.
The Beelink SER9 PRO+ Ryzen 7 H 255 has 8 Zen 4 cores, a Radeon 780M iGPU, 32 GB of LPDDR5X and a 1 TB PCIe 4.0 SSD, per its catalog listing. It is the same class of machine as the 8845HS boxes this guide originally recommended: the Radeon 780M is the iGPU the 8845HS uses (AMD), so local-LLM behavior is comparable. LPDDR5X typically runs faster than the 8845HS boxes' DDR5-5600, and generation speed on an iGPU follows memory bandwidth, so expect it to match or edge out them on the same model.
In practice that means an 8B model at Q4 (about 4.9 GB) runs at conversational but modest speed, under 20 tok/s on DDR5-5600-class bandwidth and a little higher on faster memory. That's fine for a voice assistant, coding autocomplete or a small RAG bot, and 32 GB leaves room for 14B models at reduced speed. The listing's 2.5G LAN and USB4 port make it a flexible home-lab node, and the USB4 port can host an external GPU enclosure if you outgrow the iGPU.
It replaces the GMKtec K8 Plus and Beelink SER8 from the first version of this guide. The K8 Plus listing has risen about a quarter since September, which is the opposite of a Black Friday deal, and the SER8 listing has no current price.
Who should skip it: if you think you'll want 30B-class models within a year, don't buy any 32 GB box. Save toward a 64-128 GB Strix Halo machine at the target price below.
These are still the right machines for big local models. They are not deals on November 12. The anchor to watch is launch pricing: Framework's Desktop launched in February 2025 at $1,999 for the Ryzen AI Max+ 395 with 128 GB (Framework). A 128 GB box near that price is a deal; one at nearly twice it is not, however large the percentage-off badge.
The easiest way to run 70B-class models and gpt-oss 120B on your desk, at the right price. Per AMD's spec page, the Ryzen AI Max+ 395 has 16 Zen 5 cores, a 40-compute-unit Radeon 8060S GPU and a 256-bit memory bus. At LPDDR5X-8000, that bus peaks at about 256 GB/s. On gpt-oss 120B, which activates about 5.1B parameters per token, a Strix Halo GMKtec EVO-X2 generated 47.49 tok/s in llama.cpp discussion #16578, and an August 2026 build reached 49.0 tok/s per See Hiong's write-up. The GTR9 Pro adds dual 10GbE, Wi-Fi 7 and two USB4 ports. Wait for: a 128 GB listing near the $1,999 launch-price class.
The 128 GB Strix Halo box with the most room to expand: a 320 W internal power supply, USB4 v2 (80 Gbps), dual 10GbE and a PCIe x16 slot. Check the slot's electrical lane count and card-size limits before you plan around it. Token generation is bandwidth-bound, so on the same model it lands in the same range as the GTR9 Pro; choose it for expansion and cooling headroom. Wait for: the same $1,999 launch-price class.
The box behind the 47.49 tok/s gpt-oss 120B result above, and the Strix Halo machine our readers click most. Same chip and memory tier as the GTR9 Pro. Wait for: the same $1,999 launch-price class.
One box for your storage and your 27-32B models: the same chip in a five-bay NAS chassis with dual 10GbE and five M.2 slots, diskless. With 64 GB it can't load gpt-oss 120B or 70B-class models. Its current listing sits above what 128 GB Strix Halo boxes launched at, for half the memory. Wait for: a price below the $1,999 128 GB anchor, or buy a 128 GB box instead if both are on sale.
DIY alternative. If you mainly run 7-14B models and want them fast, a desktop built around the GIGABYTE RTX 3060 12GB (Renewed) beats every iGPU here on speed. The RTX 3060 12GB's 7-9B median is 59.5 tok/s across 41 published runs (SpecPicks RTX 3060 benchmarks). Our Black Friday GPU deals guide has the other cards that clear our deal bar this week, and our 12GB-VRAM local LLM build guide has the full parts list.
Measured figures are linked to their source. Rows marked "ceiling" are theoretical bandwidth limits (peak bandwidth ÷ weights read per token), not measurements. Real runs land below them.
| Platform | Model | Gen speed | Type | Source |
|---|---|---|---|---|
| Ryzen AI Max+ 395 (128 GB) | gpt-oss 120B MXFP4 | 47.49 tok/s | Measured | llama.cpp #16578 |
| Ryzen AI Max+ 395 (128 GB) | gpt-oss 120B MXFP4 | 49.0 tok/s | Measured | See Hiong |
| Ryzen AI Max+ 395 (128 GB) | Llama 3 70B Q4_K_M (about 43 GB) | about 6 tok/s | Ceiling | 256 GB/s per AMD |
| Ryzen 7 8845HS (32 GB) | 8B Q4_K_M (about 4.9 GB) | about 18 tok/s | Ceiling | 89.6 GB/s from AMD's DDR5-5600 spec |
| RTX 3060 12GB desktop | 7-9B median | 59.5 tok/s | Measured (41 runs) | SpecPicks benchmarks |
The table makes one point clearly: Strix Halo is a capacity play. On mixture-of-experts models like gpt-oss 120B, it's fast. On dense 70B models, it's usable but slow. On small models, a discrete GPU wins.
It's soldered on every pick here, so buy the tier you'll need in two years, not today. 128 GB is the only tier that runs gpt-oss 120B and 70B-class models.
Bandwidth sets generation speed. Strix Halo's 256-bit bus (about 256 GB/s) is roughly three times a dual-channel DDR5-5600 laptop chip (89.6 GB/s: 2 channels × 8 bytes × 5,600 MT/s).
llama.cpp's Vulkan backend works on both the Radeon 780M and the Radeon 8060S without extra drivers. ROCm support for Strix Halo improved through 2026 and can add speed on Linux. Whichever you use, set the BIOS GPU memory allocation high enough for your model, or llama.cpp falls back to the CPU.
Generation pins the chip for minutes at a time. Larger chassis and internal power supplies, like the MS-S1 MAX's, hold sustained power better than the smallest boxes.
An external GPU port gives a small-memory box an upgrade path. Treat it as a bonus, not a plan, since eGPU enclosures add cost and bandwidth limits.
An always-on inference box spends most of its life idle. Mobile-class APUs idle far below a desktop with a discrete GPU, which adds up over a year of 24/7 uptime.
Every inline link above is listed in Sources.
This piece is editorial synthesis based on publicly available information. No independent first-party benchmarking is reported.
— Mike Perry
Amazon & eBay listings, plus full specs and alternatives on each product page.
As an Amazon Associate, SpecPicks earns from qualifying purchases; we also earn on qualifying eBay purchases via the eBay Partner Network. Prices shown were last tracked at crawl time and may vary — check the listing for the current price.