LLM Post-Training Tech Demo

Making a 1B model do exactly what we intend.

Iko, our AI receptionist, is a tech demo: a small LLM with about one billion parameters (1B), trained into a single role through data synthesis, post-training, evaluation, and quantization, running entirely on iPhone and iPad. The same methods scale to larger models.

Try the app (free on the App Store)

Iko, the AI receptionist of EqualFrontiers

Iko answers questions about EqualFrontiers — the company, its businesses, and its research. You can talk to Iko in Japanese or English, and Iko replies in the language you use.

Chat screen: Iko explains what EqualFrontiers does and what kind of AI it builds

When asked the same question over and over, Iko's face turns red and grumpy

Runs entirely on the device

After a one-time model download (about 0.8 GB), you can chat without a connection. Your conversations never leave your device.

Feelings in color and expression

An emotion model reads Iko's replies, and the amoeba-like character changes color and expression. Compliments cheer Iko up; pestering makes Iko a little grumpy.

Remembers long conversations

Iko keeps about 20 turns of conversation in mind. Conversations are saved as threads on the device, so you can pick them up later.

See how it works

An illustrated guide in the app explains how a small AI was raised: data synthesis, training, and quantization.

With larger models, you can go much further.

Even a small model, trained for a narrow purpose, can be made to behave this precisely. Post-train an open model with more than 100B parameters for your work, and design it to use tools such as search, calculation, and internal systems, and you may reach performance approaching frontier AI — without your data ever leaving.

Our CEO, Katsuki Ono, led the post-training sub-team for a 750B-parameter LLM in the University of Tokyo Matsuo Lab's medical LLM project. We apply the approach practiced with Iko — train with data, find weaknesses through evaluation, and fix them — directly to large models.

How we trained it.

1. Data synthesis

Using our website and papers as sources, open LLMs generated about 28,000 reception conversations, and another LLM checked each one for facts and tone.

2. Fine-tuning

Starting from Google's Gemma 3 1B, we trained all parameters (SFT), then added preference training that compares good and bad answers (DPO) and further training on ~20-turn conversations.

3. Quantization

The trained model was quantized to 4 bits — about 0.8 GB — small enough to run in a smartphone's memory.

4. Evaluation & improvement

We built tests that check source facts one by one and automatic graders that score the flow of conversations, then repeatedly added training data aimed squarely at the weaknesses they revealed.

5. On-device inference

Inside the app, llama.cpp runs the LLM and a Core ML emotion model reads the mood of each reply. Everything from training to inference was done on a single Mac.

Chart comparing the Iko tech demo with the untrained base model (Gemma 3 1B): factual accuracy on Japanese questions 27% to 67%; in long chats 21% to 55%; company fact check 23% to 53%; corrects false premises 6% to 56%.

The model and training data are public: model / GGUF / dataset (Hugging Face)

Your LLM, trained for your work.

EqualFrontiers supports post-training of open LLMs and custom LLM development — model selection, data design and synthetic data, training with SFT, DPO, and GRPO, evaluation design, quantization and inference optimization, and operation in closed / on-prem environments.

Models: open models from 1B up to the 750B scale Environment: your own compute (on-prem, closed network, or cloud) Pricing: quoted individually

Talk to us about a large-scale LLM project

* As an AI, Iko may make mistakes. Please check official company information on our website or via the contact form.

→ More on specialized LLMs for closed environments

Contact

For LLM post-training, custom LLM development, closed / on-prem LLM operation, or media inquiries, please use the form below. It helps to include the model scale, tasks, data, and compute environment you have in mind.