Models, apps, and tools. Every one built from scratch, every one designed to run without compromising your privacy.
A hierarchical recurrent language model that dynamically adjusts reasoning steps during inference for more efficient and adaptive compute usage. Built around a custom Hierarchical Reasoning Module with GQA, RoPE, and SwiGLU.
Flagship language model with Rotary Positional Embeddings, Grouped-Query Attention, and SwiGLU activation. 300M parameters, custom training infrastructure, built entirely in-house.
Compact transformer model for detecting harmful and unsafe text prompts. Designed for real-world deployment, runs fully offline without any cloud dependencies.
A tiny text encoder built on the CosmicFish architecture. Used as the language backbone inside CosmicVAE, the upcoming image generation model.
A native iOS chat application powered by CosmicFish. Talk to an AI that runs on-device, privately, without sending your data to any server.
A private messenger built on device IDs, not phone numbers. All messages are end-to-end encrypted and never stored outside your device. No account, no data, no trace.
An AI health companion powered by Gemma 4, built in Flutter for Android. Ask health questions, track symptoms, and get private on-device guidance without sharing sensitive data.
Practice job interviews with an AI voice agent. Get realistic interview simulations, instant feedback, and build confidence before the real thing.
The custom tokenizer used in CosmicFish-pico. Built in-house to give full control over vocabulary, encoding strategy, and training data coverage.
Custom eval suite generator for language models. Describe your model, get a targeted benchmark, score your outputs, and compare architectures and training runs. Built for ML engineers who build small models from scratch.