ChinaPC / AMD Ryzen AI Max+

Your model.
More memory.
Local control.

AMD Ryzen AI Max+ 395 / 495 mini workstations for memory-constrained local AI workloads. Compare 64GB, 128GB and 192GB systems, check your software path and choose around your model.

Compare configurationsWill my model fit?

395 from $2,099.00 · 64GB RAM, no SSD
495 from $3,799.99 · 128GB RAM, no SSD

FEVM FAEX1 Ryzen AI Max+ 395 mini workstation
FEVM FAEX1 · 64GB + 1TB SSD$2,149.00$50.99 below the listed GMKtec EVO-X2 64GB + 1TB price*
For local AI builders, researchers & private deployments

When the model is the reason
you are buying the machine.

Start with the constraint you are trying to solve. Memory capacity, software support and sustained throughput matter more than an AI badge.

01 / MEMORY CAPACITY

Your model has outgrown your GPU.

Consider a larger shared CPU/GPU memory pool when your current graphics-memory budget is the limiting factor. Use the model-weight table below to assess the next capacity tier.

Plan the memory budget →
02 / PRIVATE INFERENCE

Your documents need to stay local.

Build a local assistant or retrieval workflow over internal documents. A correctly configured local runtime can process prompts and files on your machine after the required downloads.

Check the software path →
03 / A DEDICATED AI HOST

You need a compact development node.

Run model experiments or a local API alongside development tools. Select CPU/GPU memory allocation, storage and networking for the workload you intend to host.

Compare the systems →
Size the workload before the hardware

Will your model fit?

Budget for model weights + KV cache + runtime buffers + the operating system. System RAM and GPU-allocatable memory are different limits.

WHY CAPACITY MATTERS70B × 4 bits ≈ 35GB

Raw weights alone exceed 24GB. This is a capacity illustration, not a speed comparison with a dedicated GPU.

24GB dedicated VRAM reference24GB
70B raw 4-bit weights~35GB
Illustrative scale · excludes every additional memory cost
Model sizeIdeal 4-bit weightsIdeal 8-bit weightsWhat to add
8B parameters~4GB~8GBKV cache for your context and parallel requests; quantization metadata; compute buffers; OS and other applications.
32B parameters~16GB~32GB
70B parameters~35GB~70GB
120B parameters~60GB~120GB

Arithmetic estimates in decimal GB: parameter count × bits ÷ 8. Actual GGUF and other quantized files can be larger. These are not model compatibility tests or total-memory requirements. For MoE models, use total stored parameters for weight capacity, not only active parameters. Context architecture, KV-cache precision and concurrency can materially change the budget.

Three capacity tiers

Buy memory for the workload.

Choose the smallest capacity that covers your complete memory budget. LPDDR5x RAM is onboard and cannot be upgraded later; SSD expansion and compatible drive types vary by system.

64GB

A lower-cost starting point

Consider for smaller quantized models, a single-user coding assistant or a modest local document workflow. Calculate the complete budget before selecting a larger model.

128GB

More room for model weights

Consider when a larger quantized model or context budget makes 64GB restrictive. On the 395 platform, AMD describes up to 96GB graphics allocation on selected 128GB systems.

192GB

Capacity for larger experiments

Consider when your calculated memory demand exceeds a 128GB system. The 495 platform supports up to 192GB system memory; AMD describes up to 160GB graphics allocation.

Capacity guidance is an estimate, not a model-fit guarantee. Graphics allocation depends on the selected system, BIOS, OS, driver and runtime. 128GB or 192GB system memory does not mean the entire capacity is GPU memory. More memory does not, by itself, increase tokens per second.

395 or 495?

395 or 495: compare the platform.

Platform specificationRyzen AI Max+ 395Ryzen AI Max+ PRO 495
CPU16C / 32T · up to 5.1GHz16C / 32T · up to 5.2GHz
Integrated graphicsRadeon 8060S · 40 CUsRadeon 8065S · 40 CUs
Platform maximum memory128GB · LPDDR5x-8000192GB · LPDDR5x-8533
Memory interface256-bit LPDDR5x256-bit LPDDR5x
NPU / total AI computeUp to 50 / 126 TOPSUp to 55 / 131 TOPS
Why choose it?Lower-cost entry to high-memory local computingMore memory headroom on selected 192GB systems

Platform maxima are not included in every configuration. The 495 specifications above refer to AMD's PRO 495 platform; confirm the exact processor and system specification on the selected product before ordering. A higher specification does not guarantee a fixed percentage speed improvement.

Find your fit

Compare systems.
Then choose the configuration.

Select a RAM and SSD configuration, then open the product page for current availability.

USD prices and availability checked November 13, 2026. No-SSD configurations require a compatible drive and operating system. Confirm shipping, applicable taxes/duties and the final total at checkout.

A clear price comparison

See where you save.

Compare the exact memory and SSD capacity. The first row compares different 395 systems; the next two compare the same model.

ConfigurationChinaPCListed reference pricePrice difference
395 · 64GB / 1TB
FEVM FAEX1 vs GMKtec EVO-X2		$2,149.00$2,199.99
GMKtec official		$50.99 lower
2.3%
GMKtec EVO-X2
395 · 64GB / 1TB		$2,159.99$2,199.99
GMKtec official		Save $40.00
1.8%
GMKtec EVO-X3
395 · 128GB / 2TB		$3,759.99$3,799.99
GMKtec official		Save $40.00
1.1%

*Public USD product-page prices checked November 13, 2026. Comparison is of listed product prices, before shipping, location-dependent taxes/duties, eligibility-based coupons or member offers. Different systems may have different cooling, ports, software and warranty terms. These are selected comparisons, not a claim of the lowest price across all retailers. 495 models are presented for capability and configuration choice.

Reference listings: GMKtec EVO-X2 · GMKtec EVO-X3

Validate the stack

The right hardware needs
the right runtime.

Check your deployment path before choosing a configuration. AMD GPU acceleration and NPU acceleration use different software paths.

Your workflowWhat to verifyStarting point
GGUF local inferenceExact model file, quantization, context, GPU offload and memory allocation. AMD documents Vulkan and ROCm runtime options in LM Studio.AMD LM Studio playbook ↗
Private document Q&A / local APIModel plus embedding runtime, document processing, data path and any network-connected plugins. Download the required files before offline operation.LM Studio offline guide ↗
PyTorch / ROCm developmentExact GPU/APU, OS, driver, ROCm release and framework version. Support for one combination is not support for all combinations.AMD support matrix ↗
CUDA-dependent softwareA working AMD-compatible backend or a different hardware path. A Radeon GPU is not a drop-in CUDA replacement.Check your application requirements
How should I compare real-world speed?

Use the same model file, quantization, prompt and context length; match runtime, driver, GPU offload, power mode and concurrency. Compare prompt processing, time to first token and sustained generation. TOPS alone does not predict LLM response speed.

ChinaPC has not published a controlled, like-for-like benchmark for these five systems. A model fitting in memory is not a guarantee of GPU residency or a target tokens-per-second rate.

Choose your next workstation

Bring your model.
Choose with a clear brief.

Shop 395 / 495 systemsRequest a configuration check
Include these details in your enquiry

Exact model or download link · quantization · runtime and OS · target context length · simultaneous users · target response speed · destination country.

We can use your brief to check configuration and order details. Include any required dispatch date or warranty question.

Processor sources: AMD Ryzen AI Max+ 395 · AMD Ryzen AI Max+ PRO 495