DFM Mimir v1: 1B-Parameter Hierarchical Reasoning Model Trained Entirely on Permissible Data, Competing With Qwen 3.5 4B

arXiv 2608.13517·medium signalXBlueskyLinkedInCopy link

Danish Foundation Models released Mimir v1, a 1-billion-parameter model on the Hierarchical Reasoning Model architecture trained from scratch on a mixture of 161 datasets using only permissible post-training data. Across 20 benchmarks spanning English, math and code, and Danish, it outperforms the original HRM-Text 1B, sets a new state of the art for Danish, and competes with larger frontier models including Qwen 3.5 4B and Gemma 4 E2B. Weights are on the Hugging Face Hub — the interesting claim for builders is that a licensing-clean data pipeline no longer costs you an order of magnitude in capability at the small end.