---
title: "DFM Mimir v1：仅用合规后训练数据实现1B参数前沿性能的开源HRM模型"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.13517v1"
permalink: "https://feed.astralor.com/items/itm_92581d00b3c6996b2f7c16c3"
published: "2026-08-13T17:37:53.000Z"
score: 75
channel: "ai"
tags: ["Danish Foundation Models","DFM Mimir v1","HRM","合规数据","开源模型"]
---

# DFM Mimir v1：仅用合规后训练数据实现1B参数前沿性能的开源HRM模型

## AI 点评

> 这篇论文值得关注，因为它证明“小而美”的模型在纯合规数据下也能与数倍大的对手掰手腕，尤其适合研究预算有限或对数据版权敏感的团队。如果你在探索低成本多语言或丹麦语场景，Mimir v1 可能比动辄几十B的模型更实用。

## 速览

DFM Mimir v1是一款仅1B参数的语言模型，基于分层推理模型（HRM）架构从头训练，仅使用合规的后训练数据。该模型在英语表现上极具竞争力，在丹麦语上创下新SOTA，并优于同体量的HRM-Text 1B，能与Qwen 3.5 4B和Gemma 4 E2B等更大模型抗衡。模型已在Hugging Face Hub开放，训练数据来自161个数据集，覆盖20个基准测试，兼顾英语、数学、代码与丹麦语。这一工作为受限于数据合规的研究者提供了高质量开源替代方案。

## 关键点

- DFM Mimir v1仅1B参数，采用HRM架构，从头训练，只使用合规后训练数据。
- 在英语、数学与代码、丹麦语共20个基准测试中，性能超越同体量HRM-Text 1B，并可与更大参数模型竞争。
- 训练数据来自161个数据集，模型已在Hugging Face Hub开放下载。

## 全文

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. The model is available on the Hugging Face Hub: https://huggingface.co/danish-foundation-models/DFM-Mimir

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_92581d00b3c6996b2f7c16c3