DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data
DFM Mimir v1 is a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture, trained from scratch using only permissible post-training data. It was trained on a mixture of 161 datasets and tested across 20 benchmarks for English, Math & Code, and Danish. The model outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B. It is available on the Hugging Face Hub at https://huggingface.co/danish-foundation-models/DFM-Mimir.
DFM Mimir v1 is a 1B-parameter open HRM model trained from scratch using only permissible post-training data. It delivers competitive English performance and sets a new state of the art for Danish, outperforming HRM-Text 1B and competing with larger models like Qwen 3.5 4B and Gemma 4 E2B across 20 benchmarks. The model is available on Hugging Face.
The model demonstrates that a 1B-parameter HRM architecture can achieve frontier-competitive performance using only permissible post-training data, suggesting that data quality and architectural efficiency can offset scale. The use of 161 datasets for post-training indicates a deliberate curation strategy. The reported performance on Danish benchmarks indicates strong multilingual transfer from a primarily English-focused training corpus.
This release lowers the barrier for researchers and organizations committed to open-source and ethically sourced data, potentially accelerating adoption of permissible-data models. It challenges the assumption that frontier performance requires massive, non-permissible datasets, and may influence data licensing and procurement strategies in the AI industry.
The model provides a cost-effective, ethically sourced alternative for organizations needing strong Danish language capabilities or seeking to avoid legal risks associated with non-permissible training data. Its open availability on Hugging Face enables commercial use without licensing fees, potentially reducing total cost of ownership for AI deployments.
Observable next signals include independent benchmark reproductions, adoption by Danish-language applications, and further releases from the Danish Foundation Models organization. If the model's performance holds up in third-party evaluations, it could spur more research into HRM architectures and permissible-data training pipelines.