AMD Ryzen AI Max+ PRO 395 in a mountable ~200 × 168 × 86 mm chassis. Up to 128GB soldered LPDDR5X, dual Thunderbolt 4, 2.5GbE, dual M.2, internal 300W PSU. Same silicon class as the sealed minis, sold as a Z workstation with ISV certification and three-year support options.
Street price (USD)
$2,300–$5,000+
Memory
up to 128GB LPDDR5X soldered
Bandwidth
~256 GB/s
Chip
Ryzen AI Max+ PRO 395
Catch
Soldered; 2.5GbE; enterprise pricing
Clip
HP Z2 Mini G1a — Strix Halo with up to 128GB, for local AI and LLM work in a mini workstation.
Andrew Marc David, Aug 20, 2025. 00:41–01:05.
What it is
HP’s Strix Halo mini workstation. Top SKU is the 16-core Ryzen AI Max+ PRO 395 with Radeon 8060S (40 RDNA 3.5 CUs), up to 128GB of soldered LPDDR5X-8000/8533 on a ~256 GB/s bus, dual M.2 (RAID capable, up to 8TB total in some configs), 2.5GbE, Wi-Fi 7, dual Thunderbolt 4 / USB4, and an internal 300W PSU. Chassis is tool-free to open, VESA/under-desk mountable, and sold with Windows 11 Pro plus HP’s enterprise stack (Wolf Security, Performance Advisor for VRAM allocation). Memory is soldered; you choose 16–128GB at order time. This page is the 128GB machine unless noted.
Who it is for
People who want a 24B–70B local model (or large MoE) in a certified mini workstation they can put on a corporate desk or mount behind a monitor, and who are fine with ROCm / llama.cpp / Vulkan instead of CUDA. Useful if you need ISV certification, longer warranty options, or the ability to allocate up to 96GB to the iGPU via BIOS/HP tools. Wrong if you need dual 10GbE, full discrete-GPU bandwidth, user-upgradeable RAM, or the absolute lowest street price for this silicon.
What it actually runs
Comfortable: 24B–32B dense at usable speed; smaller MoE models at high token rates.
Possible: 70B-class dense and larger sparse MoE (120B-class reported) when you buy 128GB. Capacity is the win; decode is still limited by the ~256 GB/s bus. Dense 70B lands nearer single-digit tokens per second on tuned Linux setups; some MoE configs report 20–40+ tok/s depending on quant and context.
Not the win: training, CUDA-only stacks, or expecting discrete-GPU token rates. Ignore vendor “125 TOPS” and “rival two high-end GPUs” slides.
Standout details
Price snapshot early Sep 2026: base and mid configs tracked from roughly $2,000–$3,000; 128GB Max+ 395 configs commonly $3,000–$5,000 street depending on storage, warranty, and channel (HP.com discounts, B&H, Amazon, eBay). List prices run higher. Stock and promotions move.
Memory is soldered LPDDR5X. Up to 96GB can be allocated exclusively to the GPU. Pick 128GB or do not buy this for large models. ECC options exist on some configs.
Networking is 2.5GbE + Wi-Fi 7, not dual 10GbE. Dual Thunderbolt 4 is the expansion path for external storage or docks.
Fans can get loud under sustained high-power load. Cooling is competent for the form factor but not silent-desktop quiet in performance mode. Tool-free access and dual M.2 are practical for a sealed-looking mini.
Honest verdict
The enterprise Strix Halo mini. Buy the 128GB config if you specifically want HP’s Z platform, certification, mount options, and support around the same silicon everyone else sells. Same bandwidth ceiling as every other Max+ 395 box. If you want dual 10GbE, look at GTR9 Pro or MS-S1 MAX. If you want repairability, get Framework. If you want CUDA, shop Spark / GX10.
Prices and stock move. Affiliate links stay off until this directory has real traffic.