Compact Local AI Workstation – Intel Core Ultra 9 285K | AMD Radeon AI PRO R9700 32GB | 64GB DDR5-6000 | 2TB NVMe | 280mm AIO | 1000W SFX Platinum
Model :
SKU : PRC-INTEL ULTRA 9 285K OE-24966
Vertical AeroDeck (Optional Accessory)
This is a compact AI workstation for running, fine-tuning and training models on your own hardware. Its core is the AMD Radeon AI PRO R9700, a professional AI graphics card with 32GB of VRAM, paired with Intel’s 24-core Core Ultra 9 285K and 64GB of DDR5. It runs large language models locally, fine-tunes them on your own data with LoRA and QLoRA, and trains custom vision and language models, all privately and without cloud GPU bills. It’s housed in the 21.3-litre Lian Li DAN B4 with a real wood front panel and cooled by a 280mm AIO.
Description
Overview: Private AI Compute, On Your Desk
Renting cloud GPUs gets expensive fast, and sending sensitive data to third-party AI services isn’t always an option. This workstation brings serious AI compute in-house.
The key spec for local AI is video memory: it determines which models you can run and train. The AMD Radeon AI PRO R9700 provides 32GB of VRAM on a professional card built for AI workloads. Behind it, the Intel Core Ultra 9 285K (24 cores) handles data preparation, preprocessing and multitasking, with 64GB of DDR5 and a 2TB NVMe SSD for datasets and model files. Everything fits in a compact 21.3L chassis that looks at home in any office.
Key Features
- AMD Radeon AI PRO R9700 32GB: RDNA 4 architecture with 64 compute units, 4,096 stream processors and 128 AI accelerators; 32GB GDDR6 on a 256-bit bus; boost up to 2,920MHz; PCIe 5.0 x16
- Intel Core Ultra 9 285K: 24 cores (8 Performance + 16 Efficient), up to 5.7GHz, 36MB cache, 13 TOPS NPU and 36 platform TOPS
- ASUS PRIME B860M-A WIFI: mATX board with a PCIe 5.0 x16 SafeSlot, PCIe 5.0 + 4.0 M.2, 4× DDR5 slots (up to 256GB), Wi-Fi 6E and 2.5G LAN
- 64GB DDR5-6000 (2×32GB): room for dataset loading, preprocessing and CPU offload of larger models
- Samsung 990 EVO Plus 2TB NVMe: up to 7,250 MB/s read for fast model loading and dataset streaming
- be quiet! Pure Loop 2 280mm AIO: 2× Pure Wings 3 140mm fans, a refill port for long-term maintenance and ARGB pump lighting
- DeepCool PS1000P 1000W SFX: Cybenetics Platinum efficiency, ATX 3.1, native 600W 12V-2×6 GPU connector, fully modular
- Lian Li DAN B4-mATX Wood Black: a 21.3L dual-orientation chassis with a real wood front panel, co-designed with DAN Cases
Performance & Technology
32GB of VRAM: the spec that matters for AI
The Radeon AI PRO R9700 is AMD’s professional card for local AI development. Its 32GB of memory holds models that won’t fit on typical 12–16GB consumer cards:
- Run larger LLMs fully on the GPU: quantised models in the 24B–32B class, such as Mistral Small, Qwen 32B and DeepSeek-R1 Distill 32B, fit in VRAM for fast, responsive inference
- Fine-tune on your own data: 32GB gives LoRA and QLoRA fine-tuning room for longer context and larger batch sizes than consumer cards allow
- Train custom models: image classification, object detection, segmentation, NLP and tabular deep learning, with room for larger batches
- Generate images and video: Stable Diffusion, SDXL and Flux pipelines in ComfyUI
- Blower-style cooling: the card exhausts hot air directly out of the case, keeping the interior cool during long training runs
Built for the AMD ROCm AI stack
The R9700 is supported by AMD ROCm, AMD’s open-source software platform for AI. It runs PyTorch, Hugging Face Transformers, vLLM, llama.cpp, Ollama, LM Studio and ComfyUI. For training and fine-tuning, we recommend Linux (Ubuntu) with ROCm, which we can install for you during the build.
A 24-core CPU for the full AI pipeline
Data cleaning, tokenisation, augmentation and preprocessing all run on the CPU. The Core Ultra 9 285K’s 24 cores keep the GPU fed with data so it isn’t left waiting. Its built-in NPU (13 TOPS) and Intel graphics can take on lighter background AI tasks at the same time.
Memory and storage for real datasets
- System memory: 64GB of DDR5 holds large datasets in memory and allows CPU offload when you want to experiment with models larger than 32GB, such as quantised 70B models, at reduced speed
- Expansion: the motherboard has four DIMM slots and supports up to 256GB, so there’s room to grow
- Storage: a 2TB Samsung 990 EVO Plus loads models and data quickly, with a free PCIe 5.0 M.2 slot and 4 SATA ports for more storage
Power and cooling for long runs
- Power: the DeepCool PS1000P delivers 1000W of Platinum-efficiency power in a compact SFX unit, with a native 12V-2×6 cable for the GPU and plenty of headroom for sustained training loads
- Cooling: the Pure Loop 2 280mm keeps the 24-core CPU cool and quiet, with large 140mm fans for low noise
Design & Ergonomics
- Wood and steel: the Lian Li DAN B4-mATX Wood Black has a real wood front panel, a refined look for offices, studios and labs
- 21.3-litre footprint: much smaller than a typical tower workstation, it fits neatly on or under a desk
- Two orientations: place it vertically (about 30% smaller footprint) or horizontally
- Designed with DAN Cases: a premium small-form-factor design with 2 pre-installed 120mm fans and flexible cooling layouts
- Quiet operation: 140mm radiator fans, a blower GPU that exhausts heat outside the case, and a Platinum-efficiency PSU keep noise down
What This Workstation Can Do
✅ Local LLM inference
- Run quantised open-source models up to the 32B class fully on the GPU (Llama, Qwen, Mistral, Gemma, DeepSeek-R1 Distill and others)
- Host a private ChatGPT-style assistant for your team with Open WebUI, Ollama or vLLM
- Build RAG pipelines over company documents, so your data stays on-premises
✅ Fine-tuning
- LoRA and QLoRA fine-tuning of 7B–14B models on your own data, such as customer support logs, internal docs, coding style or domain knowledge
- Full fine-tuning of smaller models (around 1B–3B parameters)
- Fine-tuning vision models and Stable Diffusion with LoRA or DreamBooth for custom image styles
✅ Model training
- Train computer vision models (YOLO, ResNet, ViT) for detection, classification and segmentation
- Train NLP models (BERT-class) for classification, sentiment and named-entity recognition
- Train time-series, forecasting and tabular deep-learning models for business analytics
- Use 24 CPU cores for classical ML (XGBoost, LightGBM, scikit-learn) alongside GPU work
✅ Generative AI & creative
- Image generation with Stable Diffusion, SDXL and Flux in ComfyUI
- Speech-to-text with Whisper and text-to-speech
- 3D, video editing and content-creation workloads
⚠️ Good to know
- Training scale: this is a single-GPU workstation. It’s ideal for development, fine-tuning and small-to-medium training runs. Training large foundation models from scratch still requires multi-GPU data-centre hardware.
- Software: the card uses AMD ROCm rather than NVIDIA CUDA. Mainstream AI frameworks are supported, but a few CUDA-only tools may need ROCm-compatible alternatives.
Compatibility / Use Cases
- AI developers & ML engineers: prototype, fine-tune and evaluate models locally before deploying to production
- Startups & SMEs: run a private AI assistant and custom models on company data without per-token or GPU-hour costs
- Researchers & universities: run experiments, train models and reproduce papers with full control of your environment
- Data-sensitive industries: legal, healthcare, finance and government teams that need AI without sending data off-site
- Creators: AI image and video generation, upscaling and AI-assisted content workflows
Why This Product Stands Out
- The most memory where it counts: 32GB of professional AI VRAM runs and fine-tunes models that consumer cards can’t hold, and it pays for itself compared with ongoing cloud GPU rental.
- Balanced throughout: a 24-core CPU keeps the GPU fed, 64GB of DDR5 and a 2TB NVMe SSD handle datasets, and 1000W of Platinum power and 280mm liquid cooling sustain long runs.
- A workstation you’ll want on your desk: it all fits in a 21.3L wood-fronted chassis.
- Ready to work: your AI workstation can be delivered with Ubuntu and ROCm pre-installed.
Specification
| Component | Name & Part Number |
|---|---|
| Processor | Intel Core Ultra 9 285K (OEM), 24 Cores (8P + 16E), Up to 5.7GHz, 36MB Cache, NPU 13 TOPS, LGA1851 |
| Motherboard | ASUS PRIME B860M-A WIFI mATX, Intel B860, LGA1851, PCIe 5.0, Wi-Fi 6E – 90MB1JY0-M0EAY0 |
| Graphics Card | PowerColor AMD Radeon AI PRO R9700 32GB GDDR6, PCIe 5.0, Blower – AI PRO R9700 32G-B |
| RAM | 64GB DDR5-6000 (2×32GB) |
| Storage | Samsung 990 EVO Plus 2TB PCIe 4.0 x4 / 5.0 x2 NVMe M.2 SSD – MZ-V9S2T0BW |
| Power Supply | DeepCool PS1000P 1000W SFX, Platinum, Fully Modular, ATX 3.1, Black – R-PSA00P-FE0B-JGUK |
| Cabinet | Lian Li DAN B4-mATX Wood Black, 21.3L Dual-Orientation, mATX / Mini-ITX – G99.B4X-WD.00 |
| CPU Cooler | be quiet! Pure Loop 2 280mm AIO Liquid Cooler, 2× 140mm Pure Wings 3, Black – BW018 |

































