We publish our work openly. Every paper is a step toward AI that is more efficient, more accessible, and safer for everyone.
Mistyoz AI
Large language models have achieved strong reasoning capabilities, though often at the cost of massive parameter counts and expensive inference. In this work, we explore a different direction: adaptive reasoning depth in compact language models.
We present CosmicFish-HRM, a compact language model built around a Hierarchical Reasoning Module (HRM) 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 that the model learns non-uniform reasoning behavior, allocating different numbers of reasoning steps across tasks and inputs. These findings suggest that adaptive reasoning depth may offer a promising alternative to relying solely on parameter scale for reasoning capability.
High-level and low-level reasoning cycles with learned halting.
Dynamic step allocation based on input complexity, not fixed depth.
Designed to run efficiently at small parameter counts, without sacrificing reasoning quality.
We will continue to do further research to improve AI, make it accessible, and make it safe. Every paper we publish is shared openly so the broader community can build on it.
Our research focuses on the areas where we believe the field needs more work: efficient architectures that run on consumer hardware, adaptive compute that scales with complexity, and safety systems that protect users without sending their data to the cloud.
Building models that do more with less. Compact architectures that run on your device without needing a data center.
Open weights, open research. AI that anyone can use, study, and build on, regardless of resources or geography.
Moderation and alignment research that protects users. Safety systems that work offline, without exposing private data.