開放模型
Mimir 1B Replaces Deep Stacking With Hierarchical Recurrence, Outperforms Tested 9B Models on Average Across Danish Benchmarks
The Danish Foundation Models team trained Mimir v1 from scratch, using HRM’s high- and low-level recurrent loops to approach some 4B models across the authors’ 20-benchmark evaluation. The weights are public, but “permissible data” does not mean fully open data: some corpora were used only under agreements or EU research exemptions.
The Danish Foundation Models team has released Mimir v1, a 1-billion-parameter model based on the Hierarchical Reasoning Model Text architecture. Rather than continually deepening a stack of Transformer layers, it starts with 32 layers and a 1,536-dimensional hidden state, configured with two high-level recurrent loops and three low-level recurrent loops so that the same computational modules repeatedly refine the representation. During training, truncated backpropagation limits the longest gradient path to five steps. The model uses the Gemma 4 tokenizer, a 4,096-token context window, and PrefixLM; inference currently requires FlashAttention to be configured correctly.
The training mixture covers 161 datasets, with approximately 70.48 billion tokens sampled per round. English accounts for 68.6% of the mixture and Danish for 24.7%. The team also used Gemma 4 31B to generate and review “ported data,” replacing existing instruction tasks that did not comply with its data policy. The entire model was trained on eight B200 GPUs for 1.65 million steps, with a global batch size of 262,144 tokens, in under three weeks.
Under a unified greedy-decoding setup, the authors measured Mimir at an average of 69.0 across seven English benchmarks, just below Qwen 3.5 4B’s 69.3. It averaged 64.1 across GSM8K, MATH, and HumanEval. Across ten Danish benchmarks, it achieved an average of 56.8, outperforming both similarly sized models and the tested Munin 8B and 9B series. These comparisons were conducted by the release team, however, and prompt formats, PrefixLM support, and the reasoning modes of different models are not fully equivalent. Mimir also continues to trail Gemma 4 E2B in mathematics and coding, and the authors acknowledge that its general assistant capabilities have not yet reached frontier levels.
For low-resource language engineering, the more important question is whether HRM can retain its advantages across other languages, longer contexts, and post-training with reinforcement learning. The data policy also warrants close scrutiny: some corpora supplied under agreements cannot be released publicly, while others were used under text-and-data-mining exemptions available to EU research institutions. Mimir therefore demonstrates an auditable training recipe, not a fully open corpus in which every token may be freely redistributed.