DFM Mimir v1: An Open 1B-Parameter HRM Language Model Matching Frontier Performance

Forum topic · 小凯 · 2026-08-15

Summary

Researchers introduce Mimir v1, a 1-billion-parameter language model built on the Hierarchical Reasoning Model (HRM) architecture and trained entirely from scratch using only permissible, ethically sourced post-training data. Addressing the high barrier that massive, non-permissible datasets create for open-source research, the model was trained on a mixture of 161 datasets. Across 20 benchmarks covering English, math & code, and Danish, Mimir v1 outperforms the original HRM-Text 1B and competes with larger frontier models such as Qwen 3.5 4B and Gemma 4 E2B, while setting a new state of the art for Danish. The paper (arXiv:2608.13517) is by Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, and Lukas Galke Poech, and the model is publicly available on the Hugging Face Hub.

Overview

Field: NLP Authors: Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, Lukas Galke Poech Published: 2026-08-13 arXiv: 2608.13517

Abstract

Current large language model development relies on massive, often non-permissible datasets, creating a high barrier for researchers committed to open-source and ethically sourced data. We introduce Mimir v1, a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture, that is trained from scratch and delivers highly competitive performance for English and sets a new state of the art for Danish using only permissible post-training data. Trained on a mixture of 161 datasets, Mimir v1 outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B, tested across 20 benchmarks for English, Math & Code and Danish.

Key highlights

1B parameters, HRM architecture: trained from scratch on the Hierarchical Reasoning Model architecture.
Fully permissible data: post-training uses only freely usable, ethically sourced data across a mixture of 161 datasets.
Strong results: outperforms HRM-Text 1B and is competitive with larger frontier models (Qwen 3.5 4B, Gemma 4 E2B) on 20 English, Math & Code, and Danish benchmarks.
New Danish SOTA: sets a new state of the art for Danish.
Open availability: model released on the Hugging Face Hub: https://huggingface.co/danish-foundation-models/DFM-Mimir

Resources

Paper: https://arxiv.org/abs/2608.13517
Model: https://huggingface.co/danish-foundation-models/DFM-Mimir