Tanka scales the Haiku architecture to study how increased context length, additional depth, and larger pretraining budgets affect reasoning quality and coherence in mid-scale language models.
Tanka shares its core architecture with Haiku Mini but operates at higher parameter count and longer context. The goal is straightforward: understand what changes — and what breaks — when the same design runs at a larger scale with more training data.
Tanka is an internal research model. It is not publicly accessible and is not optimized for deployment. Its role is to inform decisions about architecture and training that will eventually apply to production-oriented work.