Both chips fit a capable AI computer into a compact box, so documents never leave the office. But they suit different jobs: one is built around a large pool of fast memory for big language models, the other around lower power for smaller models and everyday office AI. Here is how to choose — and what to confirm with the factory before you pay.

The short answer
Choose a mini PC with the AMD Ryzen AI Max+ 395 and 128 GB if you need large models — big mixture-of-experts models, or dense models of 30–70 billion parameters when slower answers are acceptable — for a small team or long documents. Choose a mini PC with an Intel Core Ultra (Series 3) processor and 32–64 GB if models of 8–14 billion parameters are enough for a few users, and low power and noise matter more.
→ AI mini PC: Ryzen AI Max+ 395, 128 GB · AI mini PC with Intel Core Ultra (Series 3)
Memory: how big a model fits
A language model compressed to 4 bits needs a little more than half a gigabyte of memory per billion parameters, plus room for the conversation. As a rule of thumb, with a short context: about 6 GB for an 8-billion model, about 20 GB for 32 billion and about 40–45 GB for 70 billion. Long documents need more.
On both platforms the processor and the graphics share one pool of memory. The Ryzen AI Max+ 395 supports up to 128 GB, and AMD's Variable Graphics Memory setting can reserve up to 96 GB of it for the graphics; that memory is then not available to the processor. The Core Ultra X9 388H supports up to 96 GB. By default its graphics can use up to 57% of system memory — about 55 GB with 96 GB installed — and the Shared GPU Memory Override setting in Intel Graphics Software raises this limit; the maximum depends on how much memory is installed.
Memory soldered to the board cannot be added later, so choose the size when you order. Intel's Series 3 platform also supports replaceable LPCAMM2 memory modules; ask the factory which type the mini PC has.
→ Local AI in the office: how to choose a computer for language models
Speed: memory bandwidth matters more than TOPS
When a model writes an answer, it reads its weights from memory for every new token, so with large models the speed of an answer depends mostly on memory bandwidth, not on the TOPS figure in the brochure. The Ryzen AI Max+ 395 has a 256-bit memory bus with LPDDR5X-8000, and AMD quotes 256 GB/s for its own Ryzen AI Halo mini PC. The Core Ultra X9 388H has two memory channels and supports LPDDR5X-9600, but Intel publishes no bandwidth figure. If the bus is 128 bits wide, the theoretical peak is about 154 GB/s (9,600 MT/s × 16 bytes); the memory in a given mini PC may also run slower, so ask the factory for its actual speed. With the same large model, the AMD machine will usually generate text faster.
That is also why a dense 70-billion model is slow on both: at 4 bits it reads about 40 GB for every token, so even 256 GB/s allows at most about six tokens per second. Mixture-of-experts models are much faster, because they use only part of their parameters for each token. OpenAI's open-weight gpt-oss-120b has about 117 billion parameters, of which about 5.1 billion are active per token. AMD reports that its weights need roughly 61 GB and that it runs at up to 30 tokens per second on a laptop with the Ryzen AI Max+ 395 and 128 GB of memory, 96 GB of it given to the graphics (AMD's test, August 2025). Results on a mini PC will vary.
Graphics, NPU and software
The Ryzen AI Max+ 395 has Radeon 8060S graphics with 40 graphics cores and an NPU of up to 50 TOPS; AMD quotes up to 126 TOPS for the whole chip. The Core Ultra X9 388H has Intel Arc B390 graphics with 12 Xe cores (up to 122 TOPS) and an NPU of up to 50 TOPS; for select Core Ultra Series 3 H-series processors, Intel quotes up to 180 platform TOPS — processor, graphics and NPU combined. For language models these figures say little: memory bandwidth matters more.
For language models, the graphics does most of the work. LM Studio supports Ryzen AI Max (Strix Halo) machines through its llama.cpp runtime, and AMD supports its ROCm software on this chip under Linux; on Windows only some ROCm components are supported. For Intel, the OpenVINO toolkit supports Core Ultra Series 3 processors, Intel Arc graphics and Intel's NPU. In some versions of LM Studio, integrated graphics on the Vulkan backend are switched off by default and must be turned on. Support changes with every release, so ask the factory to confirm in writing that your software runs on the exact model.
Power, noise and size
AMD rates the Ryzen AI Max+ 395 at 55 W by default, configurable from 45 to 120 W; the mini PC maker chooses the setting. Intel rates the Core Ultra X9 388H at 25 W base power and up to 80 W maximum turbo power. A higher setting gives more speed, but needs a bigger cooler and power supply, and the fan is louder under load.
For a desk in a quiet office, ask the factory for the power setting, the noise level at full load and the rating of the power adapter. If the machine serves the whole office, plan a UPS for it.
→ Online UPS with LiFePO4 battery, 1–3 kVA · UPS for a server room or a local AI computer: which type, which battery and what size
Which one for which office
A small law firm, clinic or accounting office that wants a private assistant for its documents on an 8–14-billion model: a Core Ultra mini PC with 32–64 GB.
A team that needs 30–70-billion models, long contracts, a few users at the same time, or that develops and tests AI applications: a Ryzen AI Max+ 395 with 128 GB.
Dozens of simultaneous users, or training your own models: neither — that needs a server with professional graphics cards, which is a different budget.
What to ask the factory before you order
Get in writing: the exact processor and memory size — for Intel, dual-channel memory of at least 16 GB, which Intel requires for Arc graphics; whether the memory is soldered or a replaceable module; the power setting and the noise level at full load; the graphics memory setting, preset before shipping if you want it; the operating system and licence, or your own system image installed before shipping; confirmation that your software runs on this model; the warranty and how repairs work; the power plug for your country; and the radio test reports for Wi‑Fi and Bluetooth, which some countries require for import.
Order a sample first and check it against the specification before the batch.
→ How to order samples from China: what to check before you pay
How we help
Tell us which models you want to run and how many people will use them. We suggest a configuration, ask the factory to confirm compatibility and settings in writing, can have your system image installed before shipping, and prepare the export documents. We work with orders from US$3,000.
The specifications above were checked on the websites of AMD, Intel, OpenAI and the software developers in November 2026; chips and software change quickly, so confirm them for the exact model before you order.