DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data
- ID: a15ca3ae
- 原文链接: https://arxiv.org/abs/2608.13517
- PDF: https://arxiv.org/pdf/2608.13517v1
- 作者: Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, Lukas Galke Poech
- 日期: 2026-08-13
- 更新: N/A
- 分类: models
- 来源类型: paper
- 标签: hrm, open-data, small-model, permissible-data, danish, arxiv
- 质量评分: 4/5
- 抓取时间: 2026-08-17T04:23:51+00:00Z
中文导读
Mimir v1 是基于 HRM（分层推理模型）架构的 1B 参数语言模型，从零训练后训练数据全部来自许可可用来源（161 个数据集混合）在覆盖英语数学与代码丹麦语的 20 个基准上，它超过原版 HRM-Text 1B，并与 Qwen 3.5 4BGemma 4 E2B 等更大的前沿模型竞争，同时为丹麦语刷新 SOTA模型已在 Hugging Face Hub 开放（danish-foundation-models/DFM-Mimir）
为什么值得关注
Mimir v1 是基于 HRM（分层推理模型）架构的 1B 参数语言模型，从零训练后训练数据全部来自许可可用来源（161 个数据集混合）在覆盖英语数学与代码丹麦语的 20 个基准上，它超过原版 HRM-Text 1B，并与 Qwen 3。
收录理由：小参数量 + 分层推理架构 + 全合规数据的完整从零训练开源案例，为许可数据能逼近何种水平给出可复现参照
关键信息
- 论文标题：DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data
- 作者：Peter Schneider-Kamp, Jacob Nielsen, Gianluca Barmina, Kenneth Enevoldsen, Lukas Galke Poech
- arXiv：https://arxiv.org/abs/2608.13517
- 发布时间：2026-08-13
- arXiv 分类：cs.CL, cs.AI
- 关联标签：hrm, open-data, small-model, permissible-data, danish, arxiv
English 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. The model is available on the Hugging Face Hub: https://huggingface.co/danish-foundation-models/DFM-Mimir
English Summary
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 from a mixture of 161 datasets. It outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B across 20 benchmarks covering English, Math & Code, and Danish, setting a new state of the art for Danish. The model is released on the Hugging Face Hub (danish-foundation-models/DFM-Mimir).
Obsidian Notes
- 内容由
opencli arxiv paper拉取 arXiv 元数据与摘要生成。 - 中文导读与价值判断锚定于条目与论文摘要，未补充摘要之外的内容。