The true leader of public
AI solution
SAI-32B
On-Premises Server Series
SAI-671B
671 billion parameters
SAI-70B
70 billion parameters
SAI-32B
32 billion parameters
Video Memory
768GB
Computing Power
FP8 10560 TFLOPS
Memory
1TB
Throughput
(Total tokens/s)
1400
Concurrent Users
(Simulaneous Acccess)
100
Industrial Level MoE Architecture
Efficient inference and low computational cost
92%
OpenAi Chat GPT-4
Video Memory
192GB
Computing Power
FP16 1320 TFLOPS
Memory
512GB
Throughput
(Total tokens/s)
2100
Concurrent Users
(Simulaneous Acccess)
260
Advanced Business, Agent, and Document Analysis
82%
OpenAi Chat GPT-4
Video Memory
96GB
Computing Power
FP16 660 TFLOPS
Memory
256TB
Throughput
(Total tokens/s)
2100
Concurrent Users
(Simulaneous Acccess)
260
General Chatbot/RAG Server
71%
OpenAi Chat GPT-4
항목 | SAI-32B |
|---|---|
RAG/Agent 적합도 | 적절 |
문장 요약/추론
| 꽤 우수
|
실제 응답 퀄리티
| 자연스럽고 논리적
|
VRAM 요구량
| 48GB 이상 (4bit 기준)
|
속도 | 빠름 (32B 기준 빠름)
|
벤치마크 (MMLU 등)
| GPT-3.5 base 이상
|
추론 능력
| GPT-3.5 중간 수준
|
언어 이해 성능
| 중상급 |
파라미터 수
| 32B (32 billion)
|
항목 | SAI-70B |
|---|---|
RAG/Agent 적합도 | 매우 적합 |
문장 요약/추론
| 매우 우수 (추론력 뛰어남)
|
실제 응답 퀄리티
| 훨씬 자연스럽고 깊이 있음
|
VRAM 요구량
| 80GB 이상 (4bit 기준)
|
속도 | 느림 (무겁고 메모리 많음)
|
벤치마크 (MMLU 등)
| GPT-4에 근접하는 최고 성능
|
추론 능력
| GPT-4 대비 85~90% 성능
|
언어 이해 성능
| 상급 (GPT-4급 근접)
|
파라미터 수
| 70B (70 billion)
|
On-Premises Server
Model Stack
Attribute | SAI‑32B | SAI‑70B | SAI‑671B |
|---|---|---|---|
Suitability for RAG/Agent | Adequate | Highly suitable | Very highly suitable |
Summarization / reasoning | Quite good | Very good (excellent reasoning) | Excellent (exceptional reasoning)
|
Real‑world response quality | Natural and logical | Much more natural and deeper | Extremely natural and deep
|
VRAM requirement | 48 GB + (4‑bit) | 80 GB + (4‑bit) | 685 GB + (4‑bit)
|
Speed | Fast (fast for 32B) | Slow (heavy, high memory) | Slow (heavy, high memory)
|
Benchmarks (MMLU, etc.) | Above GPT‑3.5 base | Top performance close to GPT‑4 | Top performance close to GPT‑4
|
Reasoning capability | Mid‑range GPT‑3.5 level | 85–90 % of GPT‑4 performance | 86–98 % of GPT‑4, surpasses it in some cases |
Language‑understanding performance | Upper‑intermediate | Advanced (close to GPT‑4 level) | Advanced (close to or surpasses GPT‑4 in some areas)
|
Parameter count | 32B (32 billion) | 70B (70 billion) | 671B (671 billion) |
항목 | SAI-32B | SAI-70B | SAI-671B |
|---|---|---|---|
파라미터 수 | 32B (32 billion) | 70B (70 billion) | 671B (671 billion) |
언어 이해 성능 | 중상급 | 상급 (GPT-4급 근접) | 상급 (GPT-4급 근접 일부 능가) |
추론 능력 | GPT-3.5 중간 수준 | GPT-4 대비 85~90% 성능 | GPT-4 대비 86~98% 일부 능가 |
벤치마크 (MMLU 등) | GPT-3.5 base 이상 | GPT-4에 근접하는 최고 성능 | GPT-4에 근접하는 최고 성능 |
속도 | 빠름 (32B 기준 빠름) | 느림 (무겁고 메모리 많음) | 느림 (무겁고 메모리 많음) |
VRAM 요구량 | 48GB 이상 (4bit 기준) | 80GB 이상 (4bit 기준) | 685GB 이상 (4bit 기준) |
실제 응답 퀄리티 | 자연스럽고 논리적 | 훨씬 자연스럽고 깊이 있음 | 아주 자연스럽고 깊이 있음 |
문장 요약/추론 | 꽤 우수 | 매우 우수 (추론력 뛰어남) | 아주 우수 (추론력 매우 뛰어남) |
RAG/Agent 적합도 | 적절 | 매우 적합 | 아주 적합 |
항목 | SAI-671B |
|---|---|
RAG/Agent 적합도 | 아주 적합 |
문장 요약/추론
| 아주 우수 (추론력 매우 뛰어남)
|
실제 응답 퀄리티
| 아주 자연스럽고 깊이 있음
|
VRAM 요구량
| 685GB 이상 (4bit 기준)
|
속도 | 느림 (무겁고 메모리 많음)
|
벤치마크 (MMLU 등)
| GPT-4에 근접하는 최고 성능
|
추론 능력
| GPT-4 대비 86~98% 일부 능가
|
언어 이해 성능
| 상급 (GPT-4급 근접 일부 능가)
|
파라미터 수
| 671B (671 billion)
|
Item | SAI‑32B |
|---|---|
Parameter count | 32 B (32 billion) |
Language‑understanding performance | Upper‑mid tier |
Reasoning ability | Mid‑GPT‑3.5 level |
Benchmark (MMLU, etc.) | Above GPT‑3.5 base |
Speed | Fast (for a 32B model) |
VRAM requirement | 48 GB or more (4‑bit) |
Actual answer quality | Natural and logical |
Sentence summarization / Inference | Quite good |
RAG / Agent suitability | Adequate |
Item | SAI‑70B |
|---|---|
Parameter count | 70 B (70 billion) |
Language‑understanding performance | Upper tier (near GPT‑4 level) |
Reasoning ability | 85–90 % of GPT‑4’s performance |
Benchmark (MMLU, etc.) | Peak performance close to GPT‑4 |
Speed | Slow (heavy & memory‑hungry) |
VRAM requirement | 80 GB or more (4‑bit) |
Actual answer quality | Much more natural and in‑depth |
Sentence summarization / Inference | Excellent (outstanding reasoning) |
RAG / Agent suitability | Highly suitable |
Item | SAI‑671B |
|---|---|
Parameter count | 671 B (671 billion) |
Language‑understanding performance | Upper tier (near GPT‑4, sometimes superior) |
Reasoning ability | 86–98 % of GPT‑4, occasionally surpasses |
Benchmark (MMLU, etc.) | Top performance approaching GPT‑4 |
Speed | Slow (heavy & memory‑hungry) |
VRAM requirement | 685 GB or more (4‑bit) |
Actual answer quality | Very natural and in‑depth |
Sentence summarization / Inference | Outstanding (exceptional reasoning) |
RAG / Agent suitability | Extremely suitable |
Why should I install an AI server like ChatGPT on my local network?
클라우드 방식 (ChatGPT 웹) | 내 서버 방식 (SAI On-Premise) |
|---|---|
매월 구독료 지불 필요 | 초기 구매만 하면 계속 사용 가능, API 비용 없음 |
내부 기밀 데이터 전송 위험 | 외부 연결 없음 → 보안 완벽 |
사내 시스템과 연결 어려움 | ERP, 그룹웨어, 문서 서버 등과 바로 연동 가능 |
AI 성능 조절 불가 | 모델 종류 및 성능 선택 가능 (32B/70B 등) |
Cloud‑based approach (ChatGPT web) | On‑premise server approach (SAI On‑Promise) |
|---|---|
Unable to adjust AI performance | Ability to choose model type and performance (32B/70B, etc.) |
Difficult to integrate with in‑house systems | Can be directly integrated with ERP, groupware, document servers, etc. |
Risk of transmitting internal confidential data | No external connections → perfect security |
Requires monthly subscription fee | One‑time purchase enables continuous use, no API costs |
01
It is very cost effective.
· Build a high-performance LLM server for the price with an optimized performance balancing configuration
· Performance can be upgraded simply by adding parts, taking into account the expansion of AI module use.
02
As it is open source, you have complete control.
· GPT-4 is OpenAI's closed model → internal structure is unknown and relies on API
· SAI can directly modify or customize the model
· Suitable for sensitive data, internal corporate knowledge, and private LLM environments
03
"Good enough" performance
· Although GPT-4 is excellent, it is not suitable for general business
document writing, summarizing, code generation, and Q&A.
It can also support up to 32B
· In fact, in LLM benchmarks (MMLU, GSM8K, HellaSwag, etc.), SAI 70B
outperforms LLaMA2 70B, similar to or slightly ahead of Falcon 180B
· Not at all lacking for practical productivity + code processing + general
chatbot use
Why should I install an AI server like ChatGPT on my local network?
01
Requirements analysis
02
Hardware specification selection
03
Software installation
04
System configuration
05
AI module installation
06
Data input/fine‑tuning
07
Administrator training
08
logo_%EB%9D%BC%EC%9A%B4%EB%93%9C-%EB%A1%9C%EA%B3%A0.png)
설치 완료
01
요구사항분석
02
하드웨어 사양 선정
03
소프트웨어 설치
04
시스템 설정
05
AI 모듈 설치
06
데이터 입력/파인튜닝
07
관리자 교육
Why should I install an AI server like ChatGPT on my local network?
logo_%EB%9D%BC%EC%9A%B4%EB%93%9C-%EB%A1%9C%EA%B3%A0.png)
05
AI 모듈 설치
Why should I install an AI server like ChatGPT on my local network?
03
01
Software installation
ChatGPT에게 묻기
Requirements analysis
04
07
System configuration
Administrator training
02
Hardware specification selection
06
데이터 입력 /
파인튜닝
On-Premises Server
Function-specific usage suggestions
High-performance chatbot
GPT-4 level
Natural language conversation
Document summary and creation
Reports, emails,
Meeting minutes automation
Professional Q&A (RAG)
Based on internal documentation
Question and answer system
Code generation and analysis
Developer Assistant,
Code Description
Voice Chatbot
Whisper + TTS integration,
AI speaker implementation
Create image
Stable Diffusion
Generate prompt
LoRA Fine Tuning
Specialized for specific tasks
Building an in-house model
Fully offline support
Without internet
Capable of operating its own AI
On-Premise Server
Suggested applications by industry
Finance/Insurance
Offline chatbot,
Document summary,
Internal Regulations Q&A
Public/Defense
Intranet AI secretary,
Civil complaint automation,
Voice Assistant
Medical/
Pharmaceutical
EMR Summary,
Medical Q&A
Automated patient consultation
Manufacturing/
Industry
Manual Summary,
fault diagnosis,
Voice command system
Interpretation/
Translation
Real-time multilingual
Interpretation/Translation
SAI DePIN Cloud Service
SAI solutions are available at a lower cost to customers through NANODC.
Conveniently available as an innovative DePIN-based cloud service
Ministry of Science and ICT, Electronic Times
K-Digital Brands in 2025
Selected as the Data Center Sector Award winner!
NANODC was awarded the grand prize by the Ministry of Science and ICT and the Electronic Times for its active activities and contributions to the industry among the 20 most recognized brands representing the Republic of Korea in 2025, and was recognized for its marketability and business potential.
100G Switch
Node/Miner
Supra Worker
Aethir
GPU Server
SAi
GPU Server
Filecoin
Storage
Server
Smart UPS
All-in-one design
UPS, batteries, communications, monitoring, etc.
All the features in one
Ultra-high performance server configuration
Software and hardware
Transcendent server configuration
Low power consumption
Based on outstanding savings effect per capacity
Power Design
Ultra-integrated design
Outstanding in terms of capacity
Power design based on savings effects
Super easy management
With remote management and touchscreen
Lag and server status
Easily monitor and control
Easy installation
Space of 10sqm or more and 15kw
If you have power, you can make a micro-scale machine in one day.
Possibility to install decentralized data centers
ABOUT US
What does an AI consultant really do?
Learn about our comprehensive approach to AI consulting.
This is the space to introduce visitors to the business or brand. Briefly explain who's behind it, what it does and what makes it unique. Share its core values and what this site has to offer. Define the qualities and values that make it unique.
1
High-performance chatbot
Natural language conversation at the level of GPT-4
2
Document summary and creation
Automate reports, emails, and meeting minutes
3
Professional Q&A (RAG)
In-house document-based question-and-answer system
4
Code generation and analysis
Developer helper, code explanation
5
Voice Chatbot
Whisper + TTS integration,
AI speaker implementation
6
Create image
Stable Diffusion Prompt
generation
6
LoRA Fine Tuning
In-house models specialized for specific tasks
Build
6
Fully offline support
Self-operating AI without the Internet
deepseek On-Premise Server
Function-specific usage suggestions
Reference Sites
A variety of companies and organizations are applying SAI solutions.
Let's talk - schedule a free consultation
Get ahead of AI today