MSI Pro Max Edge AI+ Review: $5,499 Ryzen AI Max+ 395 Mini Workstation for Local AI

What MSI is selling​

The machine is built on the Ryzen AI Max+ 395, which has 16 Zen 5 cores and 32 threads. It has integrated Radeon 8060S graphics and 128GB of soldered LPDDR5X-8000 unified memory. MSI says the platform delivers up to 126 TOPS of total AI performance, including an XDNA 2 NPU rated at up to 50 TOPS. Up to 96GB of the memory can be allocated dynamically as graphics memory.

The review unit also had:

• A 2TB PCIe Gen 4 x4 NVMe SSD. The review reports a second M.2 slot, which is empty.
• Wi-Fi 7 and 2.5GbE.
• A 300W power supply built into the chassis, so there is no external brick.
• A front performance/eco mode switch.
• A UHS-II SD card reader.
• USB4 ports on the front and the back.

MSI's US store lists the unit as "Copilot+ Ready", with a one-year warranty. The review did not test the Copilot+ designation.

The review gives the size as 197.5 × 188.4 × 94.5mm and the weight as 2.9kg. MSI's store lists 3.84 × 3.94 × 9.78 inches and 6.77 lb, which matches after unit conversion. It is bigger than a typical mini PC. Jennings puts it somewhere between a mini PC and a mini tower.

Price​

TechRadar lists the price as $5,499 in the US, €5,299 and £5,999. The machine is sold through MSI and other retailers. One European outlet reported a starting price of 4,799 euros at launch, so regional pricing and configurations appear to vary. Check current listings before you buy.

What the reviewer did with it​

Jennings says Windows setup took about 10 to 20 minutes. He then ran a mix of conventional and AI workloads.

Everyday and creative work

• Word, PowerPoint, Excel and several Edge windows ran together with no noticeable slowdown.
• Adobe Bridge scrolled through thousands of Sony A7 IV JPEG and RAW files with quick thumbnails. A batch of 50 images took a few seconds to move from Camera Raw into Photoshop.
• Premiere Pro handled a five-minute edit of 4K Canon C-Log footage cut from more than an hour of material.
• The footage came from an external Thunderbolt 4 drive. Jennings could not use his NAS for editing because of the 2.5GbE port.

Local AI

• He used Lemonade to manage the models. Qwen 2.5-VL-7B handled image recognition, Qwen3-Coder-30B-A3B handled coding, GPT-OSS-20B-NPU handled the first image-sorting pass, and Flux-2-Klein-9B handled image generation.
• Sorting 2,634 images started on GPT-OSS-20B-NPU but was finished on Qwen 2.5-VL. Some images went unrecognised. Jennings says these were mostly nondescript or badly exposed.
• Qwen3-Coder ran in Visual Studio Code through the Continue extension. It produced a ten-page static site in about an hour, but it needed extra configuration. Each file creation had to be approved by hand. The result had minor JavaScript issues and left out the supplied images.
• The server session recorded 152,937 input tokens, 24,639 output tokens and 51 requests. Jennings estimates about 200 requests would be realistic to finish the site.
• The final response generated 406 tokens at 47.85 tokens per second, with a 1.99-second time to first token.
• Flux produced four seasonal images in under a minute each, with the usual extra-limb artifacts.

These are session figures from one reviewer. They are not standardised benchmarks.

Benchmarks​

TechRadar's published scores, with no units or test settings given:

Without matched comparison systems, these numbers don't rank the machine against rivals. They are best used to sanity-check your own unit.

The NPU point​

The review shows that having an NPU doesn't mean your workload uses it. Jennings saw no NPU involvement in the Qwen 2.5-VL image-sorting run, only occasional spikes. The GPU and CPU did most of the work. To use the NPU you have to pick an NPU-compatible model, and that choice varies by task. Plan for some trial and error.

Claims versus testing​

MSI markets the system as able to run 120-billion-parameter models locally. MSI's 120B claim applies to a single system, while its larger DeepSeek example needs four PCs working together. Even for the single-system claim, practical performance depends on model precision, context size, software support and memory requirements.

TechRadar did not test either claim. Its models ranged from 7B to a 30B mixture-of-experts model, plus a 20B model. The review also reports no measured acoustics, sustained thermals or clustering results, so MSI's cooling description is unverified. Treat the 120B headline as capacity, not a promise of comfortable speed.

Privacy is similar. The review says the tests ran offline, but Jennings also used online AI to choose models. Local inference keeps prompts on the machine only if the whole toolchain stays local.

Drawbacks​

• Price. It is expensive, and Jennings notes cheaper machines can match its general-purpose performance.
• Network. The 2.5GbE port is the most pointed criticism. He expected at least 5GbE at this price. If your media lives on a NAS, plan around it.
• Expansion. There is no OCuLink port for an external GPU, and the memory is soldered.
• Age. The Strix Halo chip is over a year old, which Jennings acknowledges.
• Setup time. For newcomers he suggests setting aside at least a week to experiment. He also says a from-scratch setup takes another day, and downloading the models takes about half a day.

Who it suits​

This is my inference from the reported tests, not TechRadar's claim. The machine looks best for developers, creators and small teams who want a large model and fast integrated graphics on a desk, and who will put in the setup time. That includes people moving media over USB4 or internal storage and wanting long-running jobs such as image tagging to run in the background.

It's harder to justify if you only want a fast Windows PC. It's also hard to justify if free or cheap cloud AI tiers already cover your needs. Jennings's own suggested workflow is to use paid cloud AI for knowledge and prompt-writing, and the local model for heavy lifting.

The verdict is a 4 out of 5, with 4.5 for design and performance. It is a capable machine, but it is not plug-and-play.