AI that runs inside your walls.

SJTech designs, supplies and tunes local AI for companies: the hardware, the open-weight models and, when you need one, a model trained for your work. Your prompts and data stay on machines you own.

Email us about your setup See what we do

From the maker of SuperGemma, the most-downloaded uncensored local LLM line on Hugging Face, with 200K+ downloads.

With a cloud API, every prompt and file crosses your perimeter to servers you don't control. With local AI, the model sits inside your network. Nothing crosses the line.

Three reasons companies are bringing AI in-house

Open-weight models have caught up enough that running them yourself is now a practical choice. These are the reasons our clients give.

Token bills grow with every use. Hardware is paid for once.

A cloud API charges by the token. More staff, more agents and longer contexts all mean a larger invoice every month, and the bill is hard to forecast.

A local model runs on equipment you own. After the purchase, the marginal cost of another million tokens is electricity. Enterprises are already moving: open-weight models carry a growing share of their token volume.

Open-weight models' share of enterprise AI token usage

23%A year earlier

34%September 2026

41%Expected in 12 months

Enterprise Technology Research survey of 200 enterprises, September 2026 [1]

69%of companies running open-weight models in production cite cost savings [1]

93%of enterprises have moved AI workloads off public cloud or are evaluating it [2]

56 → 41%share running production inference mainly in public cloud, last year to this year [3]

Data that never leaves can't leak from someone else's system.

Using a hosted model means sending prompts, documents and code to a third party. Retention and training terms differ by provider and contract, but in every case the data has left your network.

A local model keeps all of it inside, and can run fully air-gapped. That matters more each year: breaches are larger, and staff quietly pasting company data into outside AI tools is now a leading cause.

People affected in recent large breaches in Korea

Coupang37.56M

SK Telecom23.24M

Tving19.53M

Reported by Yonhap News [5]

43%of security incidents involved shadow AI in 2026, up from 20% a year earlier [4]

$5.39Maverage cost of a breach involving shadow AI, above the $4.99M global average [4]

91%of IT leaders would keep AI on sensitive data off standard public cloud [2]

2,383intrusions reported by Korean companies in 2025, up 26.3% [6]

Security teams need a model that doesn't refuse the job.

Hosted frontier models ship with guardrails that often decline the work defenders do every day: writing a proof-of-concept for a known vulnerability, taking malware apart, planning a red-team exercise.

An uncensored open-weight model on your own hardware answers those requests, and the sensitive findings never leave the building. Removing refusals without damaging model quality is our specialty. It is how SuperGemma became known.

We supply uncensored models for authorized work: testing systems you own or are contracted to assess.

Typical security tasks, by where the model runs
Task	Hosted frontier model	Local uncensored model
Proof-of-concept for a known CVE	Often refused	Answered
Malware sample analysis	Often refused	Answered
Phishing drills for staff training	Often refused	Answered
Red-team attack path planning	Often refused	Answered
Where the findings go	Provider's servers	Stay in-house

Cloud API and local AI, side by side

Neither is right for everything. This is where each one stands today.

	Cloud API	Local AI with SJTech
Cost	Charged per token. Grows with every user and agent.	Hardware bought once. After that, mostly electricity.
Your data	Sent to the provider's servers on every request.	Stays on machines inside your network.
Customization	Prompts and settings on a model you don't control.	Fine-tuned on your work. You own the weights.
Security work	Guardrails often refuse exploit and malware tasks.	Uncensored models answer authorized security requests.
Availability	Needs internet. Models and prices change on the provider's schedule.	Runs offline or air-gapped. Nothing changes unless you change it.
Raw capability	The largest frontier models still lead on the hardest reasoning.	Strong enough for most business tasks. We test on yours first.

What we do, in the order it usually happens

Take one step or all three. Most clients start with consulting and decide on the rest once the numbers are on the table.

Consulting

We look at what you run today and work out what a local setup would need to match it.
- Review of current AI use and API spend
- Model selection for your tasks and languages
- Security and compliance requirements
- Cost comparison: local versus API
Hardware supply

We size, source and install the machines, then set up the inference stack on them.
- Sizing from a single workstation to multi-GPU clusters
- Procurement and on-site installation
- Inference server, quantized builds, access control
- Air-gapped configurations
A model of your own

When an off-the-shelf model falls short, we train one for your work and hand you the weights.
- Fine-tuning and reinforcement learning on your tasks
- Refusal removal for security and research teams
- Evaluation against your own test cases
- You own the resulting weights

Hardware sized to the job

We match the machine to the model and the number of people using it at once. Three common shapes:

Team

One compact AI workstation on a desk or in a closet. Enough for a single team or a pilot.

For example, one NVIDIA DGX Spark: 128 GB unified memory, models up to 200B parameters [8]

Department

Several workstations linked into one cluster, serving larger models or more users.

For example, four linked DGX Spark units: 512 GB combined, models up to 700B parameters [8]

Company-wide

Multi-GPU servers in your rack or data center, built for many simultaneous users.

Sized during consulting from your concurrency and response-time targets

Model-size figures are NVIDIA's published specifications [8]. Actual hardware is chosen per project.

Where local AI fits best

Any organization whose data should not leave the building. These are the fields that ask us most often.

Finance and insurance: Internal document Q&A and compliance review under network-separation rules.
Healthcare: Summarizing clinical notes and records while patient data stays on site.
Public sector and defense: Fully air-gapped deployments with no outside connection.
Security teams: Malware analysis, red-team planning and log triage with models that don't refuse.
Manufacturing and R&D: Assistants over design documents and source code, with no IP leaving the company.
Legal and accounting: Contract and filing review under client confidentiality.

From the maker of SuperGemma

SJTech is led by Jun Song, the developer behind SuperGemma. The SuperGemma4 line is the most-downloaded uncensored local LLM line on Hugging Face, and the Super-Tune series now spans five model families in the formats enterprises run: GGUF, MLX and NVFP4.

Jun writes about local AI on X as @jun_song, where more than 50,000 people follow his work. SJTech is the same expertise, delivered to companies.

supergemma.ai Hugging Face X @jun_song

Model downloads: 200K+
Models released: 24
Model families: 5
Followers on X: 50K+

The Super-Tune families
Family	Base model	Formats
SuperGemma4	Gemma 4	GGUF, MLX, NVFP4
SuperGLM-5.2	GLM-5.2	NVFP4
SuperDeepseek-V4	DeepSeek-V4 Flash	Multi-GPU builds
SuperHY3	HY3	GGUF, MLX, NVFP4
SuperQwen	Qwen AgentWorld	GGUF, MLX, NVFP4

Questions we hear most

What is local AI?

Local AI means the language model runs on hardware your company owns or controls, in your office or data center, instead of being reached over the internet through a provider's API. Prompts, files and outputs stay inside your network.

Is a local model good enough to replace a cloud API?

For many business tasks, yes: document work, coding help, internal search and agents over private data. The largest frontier models still lead on the hardest reasoning. During consulting we test candidate models on your real tasks, so you decide from results instead of benchmarks.

How much does it cost compared with paying per token?

It depends on volume. Hardware is an upfront cost, then usage is close to free. Teams with steady, heavy usage typically reach break-even fastest. We calculate this from your current API bills before you buy anything.

What is an uncensored model, and why would a company want one?

An uncensored model has had its built-in refusals removed. Security teams need this for authorized work such as vulnerability research, malware analysis and red-team exercises, which hosted models frequently decline. Running it locally also keeps the findings private.

What hardware do we need?

It ranges from one workstation for a small team to a multi-GPU cluster for company-wide use. The right size depends on the model, the number of users and the response speed you need. We size it, supply it and install it.

Do you work with companies outside Korea?

Yes. We are based in Korea and work in Korean and English. Email sjtechkor@gmail.com with what you are trying to do.

Tell us what you run today

Send your current AI use, rough monthly spend and any security constraints. We reply with what a local setup would look like for you.

sjtechkor@gmail.com