DFM Mimir v1: Open 1B Parameter Model Achieves Frontier Performance with Permissible Data

DFM Mimir v1 has been introduced by researchers, featuring a language model with 1 billion parameters that utilizes the Hierarchical Reasoning Model (HRM) architecture. This model was developed from the ground up using only allowed post-training data. It achieves a new benchmark for Danish and shows impressive results in English, surpassing the previous HRM-Text 1B and rivaling larger models such as Qwen 3.5 4B and Gemma 4 E2B. Mimir v1 was trained on 161 diverse datasets and assessed across 20 benchmarks, including English, Math & Code, and Danish. It is accessible on the Hugging Face Hub, promoting open-source and ethically sourced research. The announcement appeared on arXiv, with the paper submitted on August 26, 2025, as part of the DFM initiative.