Ryzen AI Max+ 395 mini PC (Ryzen AI Halo, Strix Halo): price, review, and verdict for local AI
The best memory-per-euro ratio for local AI. We look at what matters for running LLMs locally : memory, bandwidth, speed, price—and who it's (or isn't) the right purchase for.
Updated on 09/10/2026
Where to buy it
Another available configuration:
AmazonGMKtec EVO-X2 64 GB / 1 TB (Ryzen AI Max+ 395) →Choose the capacity shown on the button based on the memory your models need. This memory is shared with the system and selected at purchase: it cannot be added later.
GMKtec: official European store (French-language site), with delivery in France averaging 5 to 8 days.
The button configuration is the one offered for purchase. On mini PCs, keep some memory available for the system and check the inference engine; macOS/MLX and CUDA are not interchangeable.
Affiliate links — BestLLMfor may earn a commission from purchases at no extra cost to you, which does not influence these independently determined recommendations. As an Amazon Associate, BestLLMfor earns from qualifying purchases.
Technical specifications
- MemoryUp to 128 GB unified memory (LPDDR5X-8000)
- Bandwidth≈ 212 GB/s
- ComputeRadeon 8060S iGPU (RDNA 3.5) + 50 TOPS NPU
- Power consumption~120 W
- Indicative price≈ 2 000 € for 64 GB and ≈ 3 300 € for 128 GB (GMKtec EVO-X2, official European store, September 2026)
- Largest model70B in Q4 (3-5 tok/s); excellent on MoE models (Qwen3-30B-A3B: 50-100 tok/s)
Who is it for?
Advantages and limitations
✓ Strengths
- 128 GB of unified memory at an affordable price
- Excellent on MoE models
- native x86, Windows, and Linux
- Several models: GMKtec EVO-X2, Minisforum MS-S1 Max, Framework Desktop
✗ Limitations
- Bandwidth < dedicated GPU (a dense 70B model remains slow)
- ROCm ecosystem less mature than CUDA
- Shared memory with the system, not expandable; usable capacity varies by model, context, and engine
Choices and limitations — review of September 12, 2026
Before choosing: shared memory and software
The Ryzen AI Max+ 395's memory is shared between the CPU and GPU: do not add RAM and VRAM together. Installed capacity and the allocatable share depend on the machine and its firmware. Check the mini-PC's exact configuration, the Vulkan/ROCm backend, and the test context before comparing throughput.
This page does not claim to include an in-house measurement for this hardware. To examine a reproducible external measurement: Strix Halo protocol and data from lhl. The results from a 128 GB machine aren’t the same as those from a 64 GB version.
Estimate my memory budget · Get started for free · Our measurements on another clearly identified machine
Strix Halo, Ryzen AI Max, EVO-X2: which is which?
Three names for the same thing, at three different levels: Strix Halo is AMD's codename for the chip, marketed under the name Ryzen AI Max / Max+ (385, 390, and 395 depending on the number of cores). The machines that bundle them subsequently bear their manufacturer's name:
| Machine | Memory | Recorded price | Note |
|---|---|---|---|
| GMKtec EVO-X2 | 64 GB | ≈ €2,000 (GMKtec EU store) | Our recommendation: covers 30B MoE models and dense models ≤ 32B |
| GMKtec EVO-X2 128 GB | 128 GB | ≈ €3,300 (GMKtec EU store) | For targeting 70B and large MoE models |
| Minisforum MS-S1 Max | 128 GB | live Amazon price on the button | Two 10 GbE ports and one PCIe x16 slot |
| BOSGAME M5 | 128 GB | unavailable (late September 2026) | — |
| Framework Desktop | up to 128 GB | variable | Modular, in high demand |
The complete list of Strix Halo machines on the market is kept up to date in our Ryzen AI Max+ 395 guide for local AI.
What are the actual speeds for LLMs?
As with any unified-memory machine, bandwidth determines everything—and it clearly favors MoE architectures, which read only a fraction of their weights at each token:
| Model | Architecture | Observed speed |
|---|---|---|
| Qwen3-Coder 30B-A3B | MoE — 3B active parameters | 50–100 tok/s: very smooth everyday use |
| Dense ≤ 32B (Q4) | Dense | Comfortable, including context headroom |
| 70B (Q4, 128 GB configuration) | Dense | 3–5 tok/s: usable, not responsive |
Le catalog filters models by available memory; for the face-off against NVIDIA, see Strix Halo vs DGX Spark.
Verdict
Frequently asked questions
Can the Ryzen AI Max+ 395 (Strix Halo) mini PC run a 70B LLM locally?
70B in Q4 (3-5 tok/s); excellent on MoE models (Qwen3-30B-A3B: 50-100 tok/s).
How much does the Ryzen AI Max+ 395 Mini-PC (Strix Halo) cost?
Budget ≈ 2 000 € for 64 GB and ≈ 3 300 € for 128 GB (GMKtec EVO-X2, official European store, September 2026) (starting at ≈ 2 000 $). Prices change quickly—check the current price through the purchase links.
Who is the Ryzen AI Max+ 395 Mini PC (Strix Halo) for?
Run large models (especially MoE) without blowing the budget.
Are Strix Halo and Ryzen AI Max+ 395 the same thing?
Yes. Strix Halo is AMD's internal codename; in stores, the chip is called Ryzen AI Max or Ryzen AI Max+ (the 385, 390, and 395 variants differ in the number of CPU/GPU cores). When a product listing mentions either name, it refers to the same 128 GB unified-memory platform.
Strix Halo or DGX Spark?
Strix Halo is worth considering if your priority is memory capacity for the price and your tools support its AMD GPU, via Vulkan or ROCm depending on the application. The DGX Spark remains useful for a CUDA or TensorRT-LLM environment. A 128 GB capacity does not make CUDA software compatible with AMD or guarantee the same throughput. Duel details: our 128 GB comparison.
64 GB or 128 GB of unified memory?
Choose 64 GB if the model, its quantization, and the desired context fit with some headroom for the system and tools. Move to 128 GB to load larger models or keep more context. This is total memory shared between the CPU and integrated GPU, not dedicated VRAM: the full capacity is not available to the model. On the GMKtec EVO-X2 and Minisforum MS-S1 Max configurations offered, the memory is integrated and cannot be added later.
Go further
Prices in euros (€) are French market prices including VAT, checked by QuelLLM. US prices differ: the Amazon buttons show the current US price.