- Services
Advertising disclosure: we earn commissions when you shop through the links below.
MSI's PRO MAX EDGE AI+ is a 4-liter mini PC built for local AI inference. It runs an AMD Ryzen AI Max+ 395 with 16 cores, Radeon 8060S graphics and an XDNA 2 NPU rated at 50 TOPS, for 126 total TOPS. Up to 128GB of LPDDR5X-8000 unified memory can be allocated as 96GB of graphics memory, which MSI says supports LLMs up to 120B parameters fully offline. It ships with Windows 11 Home or Pro and a 300W Flex power supply, and MSI lists a price of EUR 4799.
Sized for production serving of 70B–200B class models at full or lightly-quantized precision. Overkill for a homelab; right call when the workload pays for itself in token volume.
The PRO MAX EDGE AI+ is MSI's entry into the small-form-factor local inference market: a 4-liter aluminum mini PC built around AMD's Ryzen AI Max+ 395 (Strix Halo) with up to 128GB of unified LPDDR5X-8000 memory. It is not a gaming rig with an AI sticker on it. The entire design is oriented around one problem, fitting large language models into a box that sits on a desk and draws power from a standard wall outlet.
That problem is memory, not compute. Most consumer GPUs cap out at 16–32GB of VRAM, which puts anything above a 32B model at Q4 out of reach unless you accept heavy CPU offload and the token-generation collapse that follows. The PRO MAX EDGE AI+ sidesteps this by allocating up to 96GB of its 128GB unified pool as variable graphics memory, which MSI says is enough to run 120B-parameter LLMs fully offline. No cloud API, no data leaving the machine, no per-token billing.
Positioning-wise, this is prosumer-to-workstation hardware, not a data center product. It competes directly with Apple's Mac Studio (M4 Max/M3 Ultra with 128GB unified memory) and NVIDIA's DGX Spark, and indirectly with dual-GPU consumer builds. MSI lists it at EUR 4799 and status is currently announced. If your workload is "run a 70B–120B model locally at usable speed with a small footprint," this is the tier you're shopping in.
The headline numbers for inference:
The top models this device can run at 4-bit, ranked by fit and speed.
Specs not available for scoring. This product is missing VRAM or memory bandwidth data.
Also in Our Store
Docks, fast storage, cables, and a UPS sized for a local AI machine.
Generated from this product’s spec sheet. Editor reviews refine it over time.
For LLM inference, memory bandwidth is the binding constraint, not TOPS. A 256-bit LPDDR5X-8000 configuration delivers roughly 256 GB/s of theoretical bandwidth. That number determines your ceiling on token generation: a dense model of size S GB cannot exceed roughly bandwidth / S tokens per second, and real-world throughput lands well below that. The 126 TOPS figure matters far more for prefill, vision encoders, and small-model batched workloads than for autoregressive decode.
That's the honest framing for the "96GB GPU for AI" pitch. You are buying capacity, not speed. A 70B model at Q4_K_M (~40GB) will generate on the order of 5–6 tokens/second here. An RTX 5090 with 32GB will do the same model far faster, if it fits, which it doesn't at Q4 without offload. The PRO MAX EDGE AI+ trades raw throughput for the ability to hold the model at all.
With 96GB of allocatable VRAM plus system RAM headroom, this machine covers an unusually wide model range.
Comfortably, at high quantization:
The sweet spot, large dense models at 4-bit:
Large MoE models, where this hardware shines:
What doesn't fit: DeepSeek-R1 671B or Llama 3.1 405B at any practical quantization. MSI notes multiple units can be clustered for models up to ~670B, but that's a different budget and a different deployment.
Quantization guidance: for anything above 30B, Q4_K_M (or MXFP4 for gpt-oss) is the right tradeoff, quality loss is marginal and the speed gain over Q5/Q6 is real. Below 30B, step up to Q8_0 or FP16; you have the memory and the quality difference is measurable.
Multimodal and long context: the 96GB pool leaves substantial room for KV cache. Vision-language models like Qwen 2.5-VL 72B fit at Q4, and long-context work (128K tokens on a 70B model) is viable, though KV cache growth will eat into headroom fast, budget it explicitly. The XDNA 2 NPU and Radeon 8060S handle vision encoding well; the 126 TOPS platform rating is most relevant there.
Individual developers and researchers running local chatbots, RAG pipelines, or agentic workflows where a 70B-class model materially outperforms a 7B one. If you've been stuck at 8B because of VRAM limits, this is the upgrade path.
Teams deploying inference at the edgeclinics, law firms, defense contractors, or anyone with data-residency constraints. Fully offline 120B inference in a 4L chassis with a 300W PSU is a deployable appliance, not a lab experiment. The 2.5G Ethernet and Wi-Fi 7 make it easy to put on a small internal network.
Agentic workflow builders who need to run multiple models concurrently, a planner, a coder, and a vision model resident at once. 128GB of unified memory makes co-residency practical in a way that 24GB GPUs never will.
Who should look elsewhere: anyone doing pretraining or serious fine-tuning. LoRA adapters on models up to ~13B are feasible; full fine-tunes are not. And if your priority is maximum tokens/second on a model that fits in 24GB, a discrete GPU is the better buy.
vs. Apple Mac Studio (M4 Max / M3 Ultra, 128GB unified): The closest analogue. Apple's memory bandwidth is higher on M3 Ultra (~800 GB/s), which translates to meaningfully faster token generation on large models. The PRO MAX EDGE AI+ counters with native CUDA-adjacent tooling, ROCm, DirectML, and a much healthier Windows/Linux ecosystem for llama.cpp, vLLM, and Ollama, plus an upgradeable dual-NVMe design and a real NPU for vision and prefill offload. If you live in macOS and want the fastest dense-model decode, take the Mac. If you need Windows deployment, x86 tooling, or NPU acceleration, take the MSI.
vs. NVIDIA DGX Spark (GB10, 128GB unified): Similar capacity and a similar "local 100B+ inference" pitch, with NVIDIA's software stack as the differentiator. The MSI box offers a 300W Flex PSU, dual Gen4 M.2 slots, and standard x86 expansion, which is easier to slot into existing IT environments. DGX Spark wins on CUDA maturity; the PRO MAX EDGE AI+ wins on deployability and serviceability.
At EUR 4799, you're paying a premium over a 128GB Mac Studio and roughly par with DGX Spark. The deciding factors are your OS, your software stack, and whether you need an NPU. Choose this one when the answer is Windows or Linux, x86, and offline 120B inference in a chassis you can rack on a shelf.
