LLM Training Multi-GPU
- 2 or 4x RTX PRO 6000 96GB Blackwell GPU
- 32-core AMD TR-PRO 9975WX processor
- 256GB DDR5 ECC memory
- 4TB NVMe SSD storage
- 10Gb Ethernet, BMC remote management
- For LLM model training and inference
- In stock
MAQ Custom PCs
Run LLMs with Ease
gpt-oss-120b and gpt-oss-20b, Gemma, Llama 70B, and other leading open-weight language models deliver strong real-world performance at low cost. Well suited to on-device applications, local inference, or rapid fine-tuning without expensive infrastructure. The LLM AI Python development environment comes preinstalled, so as soon as your computer arrives, just plug it in and start creating or researching right away.
Configure an AI LLM Workstation Configure a Workstation
MAQ: Plug In and Start Creating
Build your own AI workstation with ease. On the MAQ website, configure top-tier hardware yourself, with fully transparent pricing. From the processor and memory to the latest NVIDIA / AMD AI GPUs, everything is customized to your needs. Even better, machines come with ComfyUI, Stable Diffusion, and various large language model (LLM) AI Python development environments preinstalled — plug in your new computer and start creating or researching right away.
Configure Your Workstation Configure a PC
After-Sales Support You Can Rely On
After-sales support means one less thing to worry about. If any system issue comes up, our expert team provides remote assistance to resolve it quickly. If your computer needs repair or an upgrade, we can also provide a loaner unit so your work continues without interruption (available to contracted customers). From pre-sale configuration advice to after-sales use and support, we're with you every step of the way, so you can focus on what matters most without worrying about the technical details.
Configure Your Workstation Configure a PCNVIDIA RTX PRO Solutions
AMD Ryzen 9000 Series Processors
WRX90: The Ultimate Expandability
- Supports up to 96 cores with the latest AMD Ryzen Threadripper PRO 9000WX processor
- Supports up to 2TB of 8-channel DDR5 ECC RDIMM memory
- 2x 10GbE Ethernet ports, 2x USB-C 40Gb ports
- 4x PCIe 5.0 M.2 SSD slots
- 7x PCIe 5.0 x16 slots
TRX50: The Second-Strongest Expandability
- Supports up to 64 cores with the AMD Ryzen Threadripper 9000X processor
- Supports up to 2TB of 4-channel DDR5 ECC RDIMM memory
- Built-in dual 10GbE networking, Wi-Fi 7
- Dual USB4 Type-C ports
- Ultimate expandability: 4x PCIe 5.0 x16 slots for multi-GPU configurations, plus 4x PCIe 5.0 x4 M.2 slots
AMD X870E: Ultimate Expandability for Personal Systems
- Supports the latest 16-core AMD Ryzen 9 9950X3D processor
- Supports up to 256GB of dual-channel DDR5 memory
- 10GbE Ethernet, USB 40Gb, Wi-Fi 7
- 4x PCIe 5.0 M.2 SSD slots
- 2x PCIe 5.0 x16 slots
A Different Path: Unified Memory.
The discrete-GPU path is capped by a single card's VRAM; Apple Silicon instead has the CPU and GPU share one pool of unified memory, all of which is available for inference. To load a 70-billion-parameter-plus model on a single machine, a Mac Studio needs neither multiple GPUs in parallel nor tensor parallelism.
The trade-off is the ecosystem: for training and heavy-volume image generation, the NVIDIA CUDA platform is still more mature. If your work is mainly local inference, private retrieval, and always-on agents, the Mac line's low noise and energy efficiency are real, practical advantages.
The Mac models on this page ship preloaded with MLX, Ollama, and llama.cpp. For the full model lineup and memory-tier comparison, see Mac Workstations.
128GB Unified Memory
A 70B-parameter-class model at 4-bit takes about 43GB, still leaving headroom for long context.
256GB Unified Memory
A 120B-parameter-class MXFP4 model takes about 60GB, leaving room for multiple models resident at once.
512GB Unified Memory
Run even larger open-weight models on a single machine.
Choose Your Own Case
MAQ offers a range of carefully selected PC cases across different price points and sizes to fit your needs — the only difference is that we treat every build with the same care as a handmade car.
Configure Your Own Workstation
Frequently Asked Questions
Which LLMs can a MAQ AI workstation run?
Depending on spec, it can run local inference for gpt-oss-20b/120b, Llama 3.3 70B, Qwen3 32B at 4-bit, Gemma 4, and more. The RTX PRO 6000 96GB model can load a quantized 120B model; the 32GB-class models (AI-Medium's Radeon AI PRO R9700, AI-Medium-Gemma's RTX PRO 4500) run 20B-class quantized inference or 13B full-precision inference smoothly.
What spec do you need to run a 70B-class LLM locally?
We recommend at least 48GB of GPU VRAM (such as an RTX PRO 5000, or multiple GPUs pooled together), 64GB+ of system RAM, and an NVMe SSD to speed up weight loading. MAQ's AI-High uses an RTX PRO 5000 48GB card, which sits right at this entry spec; for more headroom, see the AI-Highend with an RTX PRO 6000 96GB card.
What development environments come preloaded on a MAQ AI workstation?
Machines ship with Ollama, PyTorch, CUDA, vLLM, ComfyUI, and other inference/training frameworks preinstalled, along with agentic AI development tools such as Claude Code, Cursor, Codex CLI, LangGraph, and CrewAI. Depending on the model, matching LLM weight files (such as gpt-oss, Llama, or Gemma) are also preloaded — the machine is ready to use out of the box.
Does MAQ offer after-sales support?
MAQ provides remote technical support, help with the software environment, and a loaner service for contracted customers (a comparable-spec loaner during repairs) to help resolve system issues without interrupting your work. Contracted customers are covered by next-business-day on-site service anywhere in Taiwan.
For AI work, should I choose a Mac Studio or an RTX workstation?
If you mainly work in macOS with the MLX framework, or edit 4K/8K video, a Mac Studio's high unified memory (up to 512GB) and low power draw are a good fit. If you need the CUDA ecosystem (PyTorch training, ComfyUI, vLLM), an RTX workstation is the better choice.
Can I upgrade the GPU or memory later?
Yes. MAQ uses standard ATX cases and industrial-grade power supplies, so the GPU, memory, and SSD can all be upgraded later — by you or through us — avoiding the lock-in of an all-in-one machine.