Introducing Iko, a Receptionist AI That Runs Entirely on iPhone and iPad
EqualFrontiers has released Iko, a receptionist AI, for free on the App Store.
Iko is a technical demo: we trained a small LLM of about 1 billion parameters (1B) for a single role through data synthesis, post-training, evaluation and quantization, and made it run entirely on iPhone and iPad. Iko answers questions about our company, business and research in Japanese and English.
What Iko does
Runs only on your device: After a one-time model download (about 0.8 GB), you can chat offline. Your conversations never leave the device.
Shows feelings: An emotion model reads each reply, and Iko’s color and expression change.
Remembers long conversations: Iko answers with about 20 turns of context.
How we trained it
Data synthesis: Using our website and papers as sources, we generated about 28,000 receptionist conversations and checked each one with a separate LLM.
Fine-tuning: Starting from Google’s Gemma 3 1B, we ran full-parameter SFT, DPO, and additional training on long conversations.
Quantization: We quantized the model to 4 bits, about 0.8 GB.
Evaluation and improvement: We built fact-checking tests and automatic conversation scoring, found weaknesses, and added targeted training data in repeated cycles.
On-device inference: In the app, llama.cpp runs the LLM and a Core ML emotion model reads the replies. Everything from training to inference runs on a single Mac.
The model and training data are available on Hugging Face.
The same methods apply directly to larger models. We welcome inquiries about post-training LLMs for your business and running LLMs in closed or on-premise environments.