Unlocking the potential of small models.

Rootcomputer builds and studies language models typically under 7 billion parameters, along with the training pipelines, corpora, and evaluation tools they require.

Our work is primarily experimental. We design model architectures, assemble pretraining corpora, run multi-phase training pipelines, and study how architectural and data decisions affect downstream behavior, alignment, and reliability at small scale.

• Transformer architecture design and experimentation at small scale
• Pretraining corpus construction, curation, and mixing strategies
• Multi-phase training pipelines: pretraining, supervised fine-tuning, and alignment
• Behavioral stability, failure analysis, and generalization in compact models
• Foundational questions in AI cognition, alignment, and human-AI coexistence

• Epos 3 Our current flagship. Advances in web design, math, reasoning, and retrieval/web search through Kigo. 1.5B Open beta →
• Epos 2.1 Our last-generation flagship. Long-form conversation, reasoning, and retrieval/web search through Kigo. 1.5B Open beta →