跳转至

文章背景与核心概要

当前的语言模型开发往往严重依赖海量且版权存疑的数据集，这给致力于开源和合规数据研究的学者带来了很高的门槛。为了打破这一壁垒，DFM Mimir v1 应运而生。它是一个基于分层推理模型（HRM）架构的 10 亿参数（1B）语言模型，完全从头开始训练，并且仅使用由 161 个混合数据集组成的合规后训练数据。

在技术表现上，Mimir v1 不仅在英语任务上展现出极具竞争力的性能，更在丹麦语上树立了全新的技术先进性（SOTA）标准。在涵盖英语、数学与代码、丹麦语等领域的 20 个不同基准测试中，它成功超越了原版的 HRM-Text 1B，并能与 Qwen 3.5 4B、Gemma 4 E2B 等更大规模的前沿模型展开有力竞争，为开源社区提供了一条兼顾高性能与数据合规的全新路径。

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

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

arXiv: arXiv:2608.13517 [cs.CL]
Submitted: August 13, 2026
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Authors: Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, Lukas Galke Poech

arXiv: arXiv:2608.13517 [cs.CL]
Submitted: August 13, 2026
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Authors: Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, Lukas Galke Poech

📌 Summary

📌 Summary

当前的大语言模型开发通常依赖于海量且往往不具使用许可（非合规）的数据集，这给致力于开源和道德合规数据研究的学者带来了很高的门槛。Mimir v1 引入了一种基于分层推理模型（Hierarchical Reasoning Model, HRM）架构的 10 亿参数语言模型，对此问题给出了解决方案。Mimir v1 完全从头开始训练，仅使用由 161 个数据集混合而成的合规后训练数据，在英语上实现了极具竞争力的性能，并为丹麦语树立了全新的技术先进性（SOTA）标准。它在 20 个针对英语、数学与代码以及丹麦语的基准测试中表现优异，成功超越了原版的 HRM-Text 1B，并能与诸如 Qwen 3.5 4B 和 Gemma 4 E2B 等更大规模的前沿模型展开有效竞争。

Current large language model development typically relies on massive, often non-permissible datasets, which creates a high barrier for researchers committed to open-source and ethically sourced data. Mimir v1 addresses this by introducing a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture. Trained entirely from scratch using only permissible post-training data drawn from a mixture of 161 datasets, Mimir v1 achieves highly competitive performance in English and establishes a new state-of-the-art for Danish. It successfully outperforms the original HRM-Text 1B and competes effectively with larger frontier models—such as Qwen 3.5 4B and Gemma 4 E2B—across 20 distinct English, Math & Code, and Danish benchmarks.

📋 Abstract

📋 Abstract

当前的大语言模型开发依赖于海量且通常不具使用许可的数据集，这给致力于开源和道德合规数据研究的研究人员构成了很高的门槛。我们推出了 Mimir v1，这是一个基于分层推理模型（HRM）架构的 10 亿参数语言模型。该模型从头开始训练，仅使用合规的后训练数据，就在英语上实现了极具竞争力的性能，并为丹麦语树立了新的技术先进性（SOTA）标准。Mimir v1 在 161 个数据集的混合数据上进行训练，在横跨英语、数学与代码以及丹麦语的 20 个基准测试中，超越了原版的 HRM-Text 1B，并可与 Qwen 3.5 4B 和 Gemma 4 E2B 等更大规模的前沿模型相媲美。

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.

🔗 Resources & Links

📚 Citation & References

📚 Citation & References