On-Premise LLM Development & Fine-Tuning

LLMs that run on-premise —
trained for your work.

We fine-tune open LLMs on your data and your work, and run them on-premise or in closed networks — end to end.

Talk to us about an on-premise LLM

Sound familiar?

  • Confidential or personal data cannot be sent to cloud AI services
  • General-purpose AI doesn't follow your internal terminology or procedures
  • API costs keep growing, or you don't want to depend on external networks
  • You want your own LLM but lack training and evaluation know-how

Train an open LLM for your work and run it inside your environment, and you can use specialized AI without your data ever leaving.

Architecture of a closed-environment LLM: internal data (docs, records, manuals) is used to train, evaluate, and quantize a specialized LLM inside your environment (on-prem, closed network, or cloud), and your users and business apps use it inside the same environment. No data is sent to external AI services.

What we offer

Model selection

Choose the base open model and its size based on use case, accuracy, speed, license, and environment.

Data design & synthetic data

Pipelines that generate training data from your documents and have another model check facts and quality.

Fine-tuning (SFT, DPO, GRPO)

Supervised, preference, and reinforcement learning combined to teach the tone, procedures, and decision patterns of your work.

Evaluation design

Task-specific evaluation sets and automatic grading built first, then an improvement cycle that finds and fixes weaknesses.

Tool use & agents (when needed)

Train and evaluate models to call search, calculation, internal systems, databases, and APIs so they can carry out tasks.

Quantization & closed operation

Quantize models to fit your GPUs and servers, deploy where data never leaves, and support further training based on issues found in operation.

How we work

Process diagram: 1 model choice, 2 data design (synthesize and check), 3 training (SFT, DPO, GRPO), 4 evaluation (task-specific grading), 5 quantization, 6 operation. Weaknesses found in evaluation are fixed by going back to data design.

  1. Discovery: we learn about your tasks, data, compute environment, and target accuracy.
  2. Small PoC: we build the evaluation set first and test the effect with a small model and data.
  3. Full training & evaluation: we scale up training and fix the weaknesses evaluation reveals.
  4. Deployment & operation: we run the model in your environment and keep improving it.

Models: open models from 1B up to the 750B scale (Gemma, Qwen, Llama, etc.)
Environment: your own compute (on-prem, closed network, or cloud)
Pricing: quoted individually based on scale and timeline

Track record

  • Our CEO, Katsuki Ono, was sub-team leader of a 10-person post-training team (SFT/DPO) in the University of Tokyo Matsuo Lab's medical LLM project, running fine-tuning on models up to 750B parameters
  • Matsuo Lab LLM Development Competition 2025 — 1st in the preliminaries, 3rd in the finals (team leader)
  • Fine-tuned Gemini on about 50,000 conversation messages in Google's Project VOICE (iF Design Award Gold)
  • 20+ models published on Hugging Face

As a small-model example, our Iko tech demo trains a ~1B-parameter LLM into a single role and runs it entirely on iPhone and iPad.

FAQ

Can we talk to you even if we don't have GPU servers?

Yes. We work with your compute environment — on-prem, closed network, or cloud — and propose the scale of compute you need.

Can we start with only a little training data?

Yes. We build pipelines that generate training data from your documents and have another model check facts and quality, so you can start small.

Can it work with our internal systems and carry out tasks?

Yes. When needed, we train and evaluate tool use — calling internal systems, databases, and APIs — in a setup that stays inside your closed environment.

Should we use RAG or fine-tuning?

It depends on the use case. RAG suits referencing up-to-date internal information; fine-tuning suits teaching tone, procedures, and decision patterns. We often propose combining both.

Which models do you use?

We choose among open models such as Gemma, Qwen, and Llama based on your use case, scale, license, and environment.

→ See all AI services

Contact

For on-premise / closed-network LLM development or LLM fine-tuning and post-training, please use the form below. It helps to include the tasks, data, model scale, and compute environment you have in mind.