On-premise sLLM
Security-Specialized Small Large Language Model
A security-specialized sLLM solution that eliminates concerns about external data leakage. Co-sold through a strategic partnership with Jiransoft.
Last updated:
Why On-premise sLLM?
LLMs accessed via external APIs are fundamentally vulnerable to security risks. Korea's Personal Information Protection Commission confirmed that personal data of 687 Korean ChatGPT users was leaked. Samsung Group has banned employees from using external generative AI and built their own model, Samsung Gauss.
How to Prevent Data Leakage at the Root?
The key is building an environment completely isolated from external networks. By deploying an on-premise server fully separated from the internet and training a business-specific AI on your own data, you can completely block the possibility of internal data leaking outside.
SUVsoft's sLLM Solution
Our security-specialized sLLM is co-sold through a strategic partnership with Jiransoft. It operates safely within your corporate intranet via the AXworks platform, providing optimal business fit through domain-specific fine-tuning tailored to your organization's data.
Key Features
Private Network
Operates solely on internal infrastructure without internet connection, maximizing security.
Domain Tuning
Fine-tuned on company-specific documents and data for high business relevancy.
Jiran Collaboration
Reliable models proposed with top security expert Jiransoft.
Multi-Agent System
Supports multi-LLM configurations for complex process automation.
Specifications
Model and training
- Base model
- Open-weight models, no vendor lock-in
- Fine-tuning
- On your documents, manuals and policies
- Model size
- Matched to server capacity, scaled from PoC
- Multi-agent
- Complex processes split across agents
Deployment
- Model
- On-premise. Runs only on your infrastructure
- Network
- Operates with no internet connection
- Platform
- AXworks, co-sold with Jiransoft
- Hardware
- Configuration proposed after reviewing your GPU servers
Operations
- Grounding
- Answers from internal documents, with sources cited
- Permissions
- Reference scope partitioned by user permission
- Integration
- Groupware, ERP and internal wiki via API
How Deployment Works
- 1
Environment review
We check your GPU server capacity and network setup. This determines the model size you can realistically run.
- 2
Document preparation
Internal documents for training and reference are collected and organised. Whether they are current, and whether obsolete policies are mixed in, decides answer quality.
- 3
PoC
Answer quality is checked against questions that actually come up in your work, including whether sources are cited correctly and whether the system admits when it does not know.
- 4
Deployment
Built on your servers and integrated with groupware and ERP. Departmental document permissions are configured at this stage.
- 5
Operation
Question history is used to keep improving the areas where answers are weakest.
In Practice
“What is the travel expense policy?”
Asked directly instead of searching the policy manual. The source clause is shown alongside, so it can be verified.
“Summarise what we proposed to this client last time”
Finds and summarises past proposals and meeting notes. Especially useful during handovers.
“Searching internal technical documents”
Scattered design documents and issue records found in plain language, with sensitive technical material never leaving the network.
“Work with material that cannot leave the organisation”
Running disconnected from the internet means organisations that are not permitted to use external AI services can still adopt it.
Learn more about the AXworks platform, co-sold with Jiransoft.
Visit AXworks Product PageFAQ
ChatGPT sends data to external servers, but on-premise sLLM operates only within your corporate network. Sensitive internal documents, customer data, and business information can be utilized with AI without any risk of external leakage.
Yes. We fine-tune the model on your internal documents, manuals, and policies to build a company-specific AI. This significantly improves accuracy compared to general models.
Suitable for companies of all sizes, from SMEs to enterprises. Model size can be adjusted to match server capacity, and we support a phased expansion approach starting from a PoC pilot.
Yes. It can integrate with groupware, ERP, internal wikis, and other systems via API. The AXworks platform enables custom integration for various business environments.
Not ChatGPT itself. Open-weight models such as gpt-oss are deployed on your own servers and used conversationally. The experience is similar; the difference is that no data leaves your network.
It depends on model size. Smaller models can run on a single GPU; larger models need more memory. Because model size is matched to available hardware, we review the servers you already have before proposing a configuration.
Open-weight models are used, so there is no vendor lock-in. A model is chosen to suit the workload and your server capacity, and can be swapped when a better one appears.
The system is configured to answer from your internal documents and to cite the source document alongside the answer. Configuring it to say it does not know when no source is found matters more in enterprise settings than raw fluency. That threshold is agreed during deployment.
Yes. Reference scope is partitioned by user permission. Making sure restricted material such as HR or finance documents is never cited in another department's answer is a key design item during deployment.
It depends on server readiness and on how much internal documentation exists and how well organised it is. We recommend a small PoC first, to see how well real business questions are answered, and agreeing a timeline from there.
Interested in On-premise sLLM?
Leave your info and our team will reach out shortly.