Signal DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissively-Licensed Data

Summary

Current large language model development typically relies on massive, often non-permissible datasets, creating a high barrier for researchers committed to open-source and ethically sourced data practices. The paper introduces Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model architecture, trained from scratch using only permissively licensed post-training data. Despite the restrictive data-sourcing constraint and small parameter count, the model delivers highly competitive performance in English. The authors describe this as setting frontier-level results at this parameter scale. The work is positioned as a proof point that ethically sourced, openly licensed training data need not sacrifice model capability. This is relevant to the broader AI governance debate over training-data provenance.