CosmicFish · Mistyoz AI

Intelligence, reimagined

A home-grown family of compact language models, built from scratch and designed to run directly on your phone, laptop, or desktop. No cloud, no account — private by default, free to use.

CosmicChat is available

The iOS app that runs CosmicFish on your iPhone — Ceres and Eris included — entirely on-device.

Run anywhere, anytime

Optimized builds keep models fast and battery-conscious. Your conversations stay on your phone — nothing is sent to a server.

Breaking new ground in AI accessibility

CosmicFish is one of India's first home-grown language model families built from scratch to run efficiently without internet access. Designed for phones, laptops, and desktops, it prioritizes privacy, low resource usage, and fast performance — developed by Mistyoz AI.

100% offline & private

Everything runs on your hardware. Your data never leaves your device.

Free forever

Open weights released under Apache 2.0, with no usage limits or fees.

Runs anywhere

Compact by design — built to fit comfortably on everyday phones and computers.

Improving steadily

Our roadmap includes larger, specialized, and more capable models.

Open source · Apache 2.0
01
CosmicFish-Pico
The smallest model · Cosmicizer tokenizer · tooling included
19M
02
CosmicFish-90M
Compact, for everyday tasks
91M
03
CosmicFish-120M
Balanced performance and cost
121M
04
CosmicFish-300M
Flagship · strongest in code and reasoning
369M
Sample generations

See CosmicFish in action

CosmicFish-300M

Prompt: Explain gradient descent in short

Gradient descent is an optimization algorithm used to find the minimum of a function. It works by iteratively updating the parameters in the direction that most reduces the loss.

CosmicFish-120M

Prompt: How do neural networks learn?

A neural network learns through training on labeled examples. During training it adjusts its internal parameters to minimize the error between predicted and actual outputs, then generalizes to new data.

Under the hood

Advanced architecture

Cutting-edge transformer techniques, optimized for efficient on-device performance.

1
Rotary positional embeddings
2
Grouped-query attention
3
SwiGLU activation
4
RMSNorm
5
Quantization optimization
6
Comprehensive training data
Research

CosmicFish-HRM

A study of adaptive reasoning in compact models — a hierarchical recurrent mechanism that lets the model decide how much computation each input deserves.

Your data stays with you

Unlike cloud-based AI, CosmicFish runs entirely on your device. Your inputs and outputs never leave your hardware.

No internet required

Use it in sensitive environments or areas with limited connectivity — no data is sent anywhere.