CosmicFish 300M
CompactA 300M-parameter model built to punch above its weight. Small footprint and fast inference for on-device and resource-constrained deployments.
Language models
CosmicFish
A family of compact, efficient language models built from scratch by Mistyoz AI. Small enough to run on your device, smart enough to ship real products.
The family
Every size shares the same architecture and philosophy: compact, efficient, and private by design.
CosmicFish 2B
Current bestThe flagship of the family. A 2B-parameter model that balances capability and efficiency for reasoning, instruction-following, and retrieval.
CosmicFish 4B
Coming soonThe next step in the family. More capacity, the same philosophy. Compact, efficient, and private by design.
Coming soon
Why CosmicFish
Built from the ground up. Every model is architected for the real world. Devices you own, data that never leaves your control.
Compact & efficient
Every parameter earns its keep. CosmicFish models do more with less, running comfortably on hardware you already own.
Private by design
Models that run locally. Your data never leaves your device. No cloud round-trips, no telemetry, no exposure.
Open & accessible
Open weights, open research. Anyone can use, study, and build on CosmicFish, regardless of resources.
Research
Adaptive reasoning in compact language models.
Research · Paper
CosmicFish-HRM
Adaptive Reasoning via Hierarchical Recurrent Mechanisms in Compact Language Models
Large language models achieve strong reasoning, though often at the cost of massive parameter counts and expensive inference. We explore a different direction: adaptive reasoning depth in compact language models.
CosmicFish-HRM is built around a Hierarchical Reasoning Module that dynamically allocates reasoning compute during inference. Instead of applying fixed computation to every input, the model iterates through high-level and low-level reasoning cycles and learns when to halt based on input complexity.
Our results show the model learns non-uniform reasoning behavior, allocating different numbers of reasoning steps across tasks and inputs, suggesting adaptive reasoning depth as a promising alternative to relying solely on parameter scale.
