DFM Mimir v1 - An Open HRM Delivering Frontier Performance at 1B Parameters - Using Only Permissible Post-Training Data

Sep 15, 2026

Uploaded by Finn

Current large language model development often relies on massive training datasets that may not meet permissibility requirements, creating a barrier for researchers focused on open, ethically sourced data. Mimir v1 introduces a 1-billion-parameter hierarchical reasoning model trained from scratch using only permissible post-training data. Trained on 161 datasets, it delivers competitive results across English, Math & Code, and Danish benchmarks, including synthetic “transplant datasets” to replace non-permissible sources without sacrificing performance or data rights.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages23

Reading Time58 min

AccessFree to download

Introduction

Datasets

Category Distribution

Results

arXiv : 2608 . 135 17v2 [cs .CL] 20 Aug 2026 DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data Peter Schneider-Kamp 1 ,2 ,∗ Jacob Nielsen 1 ,2 , Gianluca Barmina 1 , Kenneth Enevoldsen3 , Lukas Galke Poech 1 , 1 University of Southern Denmark 2 Ordbogen A/S 3Aarhus University 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 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. The model is available on the Hugging Face Hub: [https://huggingface.co/danish-foundation-models/DFM-Mimir](https://huggingface.co/danish-foundation-models/DFM-Mimir) Date: 14 August 2026 Correspondence: Peter Schneider-Kamp via [petersk@imada.sdu.dk](petersk@imada.sdu.dk) Figure 1: Aggregate results across three categories English – Math& Code, and Danish – comparing DFM Mimir 1B against HRM-Text 1B, Qwen 3.5 2B and Gemma 4 E2B, displaying highly competitive performance across 20 benchmarks. For Danish, we also compare against three Munin variants post-trained for Danish from Apertus 8B, Qwen 9B, and Mistral 9B. 1 Introduction In recent years, Large Language Models (LLMs) have catalyzed a paradigm shift in artificial intelligence, characterized by rapid iterations and significant advancements in emergent capabilities. However, as noted by Wang et al. [2026], current development is largely driven by a “monolithic recipe” consisting of massive, multi-stage pipelines and training on exorbitant volumes of data. This approach not only necessitates vast computational resources for pre-training but also creates a prohibitive entry barrier for the broader community of researchers and practitioners. For national initiatives such as the Danish Foundation Models project 1 [ Enevoldsen et al. , 2023], which adheres to a philosophy of using exclusively permissible and, whenever possible, openly licensed data, training a capable LLM from scratch is often infeasible given the limited pool of high-quality data for a language such as Danish. Consequently, it has been challenging to provide fully permissible base models on LLM platforms to serve as foundations for post-training objectives. To address these constraints, we employ the HRM-Text framework, which enables focusing on post-training data during the initial training phase, thereby facilitating the creation of a viable base model for the wider community. In this technical report, we present Mimir v1 , a 1-billionparameter hierarchical reasoning model (HRM) trained from scratch, utilizing the architecture proposed by Wang et al. [2026] . Mimir v1 is optimized for Danish and English tasks and has undergone instruction-tuning on a curated mixture of 161 datasets, comprising approximately 70.5 billion tokens per epoch. Furthermore, as certain datasets used in [Wang et al. , 2026] do not align with DFM’s permissibility standards, we demonstrate the efficacy of synthetically generating “transplant datasets”, replacing non-permissible data with synthetically generated permissible variants. Our results indicate that these synthetic alternatives achieve comparable or superior performance without compromising data rights, further underscoring the relevance of the HRM approach for low-resource linguistic domains, empowering communities of practitioners and researchers with small, capable and fully permissible models, with both low training and inference require

FAQs

What problem does Mimir v1 address in large language model development?

It addresses the reliance on non-permissible datasets that blocks researchers committed to open and ethically sourced data from training capable models from scratch.

How is Mimir v1 trained and what does it use as data sources?

Mimir v1 is trained from scratch as a hierarchical reasoning model, using instruction tuning on a curated mixture of 161 datasets totaling about 70.5B tokens per epoch, drawn from permissible sources on the Hugging Face Hub.

How does Mimir v1 handle datasets that do not meet permissibility standards?

It replaces non-permissible data by synthetically generating “transplant datasets,” producing permissible variants while maintaining comparable or superior performance.

Sep 15, 2026

Uploaded by Finn