01 / Build referenceCOMPACT INTELLIGENCE
AI mini PC
A compact starting point for local AI and personal research. Plan model size, memory and expansion together.
Discuss this buildENGINEERED FOR THE REAL WORLD
From an idea.
To a working system.
AI mini PCs, graphics workstations, custom servers and Korean IDC. Hardware and software designed around the work you do.
One machine. One model. A whole new infrastructure.
STAEX / SYSTEM ENGINEERINGBuilt to
work.
work.
SYSTEMS / BUILT AROUND YOU
Find the shape
of your next system.
From compact personal research to team compute and data operations. Start with your workload; define the build together.
01 / Build reference
02 / Build referenceCREATIVE COMPUTE
GPU workstation
A desktop for AI development, 3D rendering and video work. Select graphics, cooling and power around your workload.
Discuss this build
03 / Build referenceRACK-READY SYSTEMS
Custom server
Compute and storage systems built around your requirements. Review GPU spacing, chassis, power and IDC installation.
Discuss this build
04 / Build referenceYOUR DATA, YOUR SPACE
NAS & backup
Private storage for sharing, archiving and independent backups. Design disks, access control and recovery as one system.
Discuss this buildPhotography illustrates equipment formats and environments. Actual models, specifications, prices and lead times are confirmed per project.
BUILD YOUR SYSTEM / REQUEST A QUOTE
Configure first. Quote next.
Choose a system and its hardware, or start with a model and token targets. An engineer will confirm compatibility, supply and pricing.
These are enquiry options. CPU, motherboard, memory type, slots, power, cooling and GPU edition require confirmation. Options do not imply stock or a validated build.
Sizing method & limitations
Generic dense-model assumptions: 8B/14B/32B/70B/120B use 32/48/64/80/96 layers, 8 KV heads, head dimension 128 and an FP16 KV cache. Weights = parameters × bits/8 plus 12% metadata. Runtime overhead is at least 2 GiB, with 15% overall headroom. Context and concurrency increase KV memory. Mini sizing budgets up to 75% of shared RAM; discrete GPUs budget 85% of nominal total VRAM. This does not certify that a model will run.
Speed controls intake priority only: Mini starts at ≤30 tokens/s per stream and ≤2 streams. Above 60 or 2 streams favours professional systems; above 120 or 8 streams favours IDC. Additional GPUs are planned after groups of 4 streams. These are routing heuristics, not benchmarks. MoE, vision, training and fine-tuning need a separate review.
GPU memory is not automatically pooled. Splitting a model needs compatible sharding, runtime and interconnect. GPU count does not imply proportional speed.
FROM DESKTOP TO DATA CENTER
Beyond the hardware.
Into your environment.
Connect equipment selection, software development and IDC preparation in a single project.
01 / CONNECTED ENGINEERING
Every layer.
One build partner.
From choosing components to bringing an environment online. A configuration shaped around your workload and budget.
/ 01
AI compute
Compute that fits the workload.
Plan GPU configurations and environments for model training, inference and rendering.
- GPU configuration planning
- Training & inference
- Dedicated compute
/ 02
Custom servers
Workload first. Components second.
Component selection, assembly, thermal and power planning, and validation for GPU and storage servers.
- GPU & storage servers
- Assembly & validation
- Power & thermal design
/ 03
Personal systems
Private infrastructure, on your desk.
Dedicated workstations and compact systems for research, creative work and local AI.
- AI workstations
- Creator systems
- Compact & quiet builds
/ 04
Private NAS
A home for data you control.
NAS design and assembly for file sharing, backups, remote access and future expansion.
- Drive & RAID planning
- Access & permissions
- Independent backup design
/ 05
Korea IDC colocation
A place for your systems to run.
Coordinate rack space, power, network requirements and installation across preferred regions in Korea.
- Region & rack planning
- Power & connectivity
- Installation & remote work scope
/ 06
AI software
Turn a model into a working tool.
Private LLM development, video AI pipelines, model integrations and optimisation software.
- Private LLMs & RAG
- Video AI workflows
- Model & runtime optimisation
02 / YOUR MODELS. YOUR WORKFLOW.
Beyond deployment.
Built around your work.
Connect model selection, data flows, inference engines and user tools into one considered system.
PRIVATE INTELLIGENCE
A language model connected to your knowledge.
Assess Kimi and other models whose licences permit dedicated deployment. Design retrieval, access control and a tailored interface around your data.
- Model selection and hardware sizing
- Internal search, RAG and tool integrations
- Permitted fine-tuning and evaluation
STAEX / SYSTEM BLUEPRINT01
01Your data
02Search / RAG
03Private LLM
04Your tools
01model / Kimi & open-weight LLMs
02data / your knowledge base
03access / controlled by design
DESIGNED AROUND YOUR WORKLOAD
Scope depends on model licences and hardware requirements. A MiniMax video API integration does not imply local deployment of its video model weights. Product names belong to their respective owners; no official affiliation is implied.
03 / INFRASTRUCTURE IN KOREA
Korean infrastructure.
Your operating requirements.
Review location, rack space, power and connectivity together. Connect custom system assembly and IDC installation in one project.
Preferred region
Regions are enquiry preferences. The facility, availability, connectivity and installation terms are confirmed in the proposal.
Discuss this regionKOREA / LOCATION PLANNING
PREFERRED REGIONSeoul / Capital areaKR / INFRASTRUCTURE
01
Rack space
Units, shared racks and dedicated racks
02
Power planning
Equipment load and power configuration
03
Connectivity
Bandwidth, circuits and IP requirements
04
Installation
Delivery, cabling and remote work scope
04 / FROM BRIEF TO BUILD
A clear path
from need to operation.
- 01
Define
Understand your goals, data, environment, budget and timeline.
- 02
Design
Propose hardware, software and IDC scope with a tailored quotation.
- 03
Build & validate
Assemble, develop, deploy and test against agreed criteria.
- 04
Hand over
Deliver documentation and validation results, and agree ongoing support.
Before we build.
Can you work with existing equipment?
We review its specifications, condition and expansion options, then propose what can be reused.
Can every model run on private infrastructure?
No. We distinguish dedicated deployment from API integration according to licences and available weights, and review any data sent outside your environment.
Does RAID replace a backup?
RAID does not replace an independent backup. We assess separate storage, systems or locations against your recovery requirements.
Do I pay on this website?
Submit an enquiry to confirm scope and receive a quotation first. The website does not charge you or reserve IDC capacity automatically.
LET’S BUILD WHAT’S NEXT.
What would you
like to build?
Start with your goals or your current environment. We will use your project requirements to plan the next step.
BASED IN GIMPO, KOREAsta express Inc. · 주식회사 스타익스프레스Connecting hardware, software and IDC.