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, 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, comprising approximately 70.479 billion tokens per epoch. Mimir 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.
Evaluation
Mimir is evaluated across Danish, English, and Math & Code benchmarks. The figure below shows average performance by subject area across compared models.
Average scores by subject area
English benchmark results (Best scores in bold.)
English benchmark results (full datasets).
Model
BoolQ (Acc)
Winogrande (Acc)
Hellaswag (Acc)
MMLU (Acc)
ARC-C (Acc)
DROP (F1)
GovRep. (R1)
Avg.
~1B models
Mimir 1B
87.8
73.5
67.3
57.5
81.6
83.1
32.0
69.0
HRM-Text 1B
87.5
70.4
60.4
58.7
82.2
78.1
25.4
66.1
Qwen 3.5 0.8B
69.8
48.9
37.0
51.5
68.4
45.2
32.5
50.5
Gemma 3 1B
62.4
49.1
30.6
37.5
43.5
7.0
29.5
37.1
OLMo 2 1B
67.2
51.0
42.4
41.6
48.1
12.4
37.7
42.9
2--3B models
Qwen 3.5 2B
80.8
53.4
64.6
62.8
82.7
31.3
31.5
58.2
SmolLM3 3B
84.3
60.3
65.1
60.2
79.5
54.0
38.1
63.1
4--5B models
Qwen 3.5 4B
87.0
70.0
83.2
75.8
92.9
48.0
27.9
69.3
Gemma 4 E2B
64.1
56.7
55.6
59.3
69.8
57.3
33.6
56.6
Gemma 4 E2B (think)
83.4
63.0
55.8
72.0
86.8
70.8
34.7
66.6
Math & Code benchmark results (Best scores in bold.)
Model
GSM8K (Acc)
MATH (Acc)
HumanEval (Acc)
Avg.
~1B models
Mimir 1B
89.9
45.8
56.7
64.1
HRM-Text 1B
84.8
56.0
0.0
46.9
Qwen 3.5 0.8B
49.1
36.1
30.5
38.6
Gemma 3 1B
49.7
37.2
42.7
43.2
OLMo 2 1B
59.4
18.8
15.9
31.4
2--3B models
Qwen 3.5 2B
73.7
55.7
47.6
59.0
SmolLM3 3B
80.0
62.2
61.6
67.9
4--5B models
Qwen 3.5 4B
60.5
56.5
78.0
65.0
Gemma 4 E2B
88.3
64.2
73.8
75.4
Gemma 4 E2B (think)
90.3
49.1
72.0
70.5
Danish benchmark results (Best scores in bold.)
Model
Angry Tweets (Acc)
DaLA (F1)
GEC (EM)
PIQA (Acc)
Daisy (EM)
WikiQA (EM)
WMT (chrF)
N.News (chrF)
IFEval (Acc)
Hellaswag-DA (Acc)
Avg.
~1B models
Mimir 1B
67.4
96.1
85.6
53.7
9.6
66.8
53.9
35.87
63.9
35.3
56.8
HRM-Text 1B
42.4
26.7
0.5
13.0
0.0
34.9
25.4
26.76
18.5
28.8
21.7
Qwen 3.5 0.8B
53.8
51.0
0.7
56.5
0.7
41.6
37.8
35.30
39.6
25.0
34.2
Gemma 3 1B
54.4
41.0
3.3
72.2
1.4
42.6
45.1
35.56
47.2
24.8
36.8
OLMo 2 1B
33.6
48.7
0.2
75.0
0.0
8.4
30.0
33.77
32.5
26.7
28.9
2--3B models
Qwen 3.5 2B
61.6
36.4
8.0
25.0
2.5
49.4
45.6
34.85
56.1
24.7
34.4
SmolLM3 3B
63.2
33.5
3.3
51.9
2.2
0.3
37.3
35.98
49.8
40.1
31.7
4--5B models
Qwen 3.5 4B
69.1
50.1
42.6
70.4
4.7
57.1
52.1
37.03
73.7
34.7
49.2
Gemma 4 E2B
64.6
56.7
36.9
46.3
5.6
44.1
55.2
35.67
75.5
25.6
44.6
Gemma 4 E2B (think)
67.7
66.8
23.4
63.9
5.1
59.3
56.0
36.30
81.2
39.0
49.9
8--9B models
Munin-Apertus 8B
60.6
46.1
42.1
81.5
12.5
49.9
55.8
30.30
53.0
24.5
45.6
Munin-Mistral 8B
61.3
48.8
26.4
76.9
8.4
48.4
51.8
32.92
67.8
33.6
45.6
Munin-Qwen 9B
69.1
60.6
11.4
38.9
5.4
55.7
56.1
35.89
71.8
34.3
43.9
Model details
Architecture
Parameters
Hidden size
Layers
Attention heads
Vocab size
Context length
Training steps
Tokens per epoch
License
HRM-Text
~1B
1,536
16
12
262,144
4,096
1,750,000
~70.5B
Apache 2.0
Technical Report
Training was performed using a fork of HRM-Text. Further details are provided in our technical report here
Usage instructions
When you run Mimir, please make sure to follow these instructions to ensure that the model is not operating out of distribution.
Use the Gemma4 tokenizer and chat template
Use a PrefixLM attention mask
Memorisation audit
We run two audits independently from each other across four data categories: synthetic instruction-tuning data derived from agreement-backed sources (A), instruction-tuning data from Hugging Face sources with uncertain opt-out status (B), instruction-tuning data from Hugging Face sources with high confidence of no applicable opt-out (C), and other low-risk synthetic and reasoning post-training datasets (D). All remaining datasets are either covered by open licenses, in the public domain, or synthetic datasets derived from such openly licensed or public domain datasets. Our memorisation audit thus covers all data sources where memorisation risk is meaningful. Full details are provided in the appendix of the technical report. Targeted prefix attacks yielded predominantly short
matches, with only 0.00021%–0.00073% of training documents have been retrieved with 50 tokens or more across categories. Across all model-input evaluations in the second audit, only 0.000044% corresponded to coherent prose and 0.00000073% to expressive prose, with no high-priority copyright findings, indicating a generally low copyright-related memorisation risk under the tested scenarios.
Limitations
Mimir v1 was trained on Danish and English data only. It will likely have poor performance on other languages. The model has not been specifically aligned for safety and may reflect social biases present in its training data.
License
This model is released under the Apache License 2.0.
1@misc{schneiderkamp2026dfmmimirv1open,
2 title={DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data},
3 author={Peter Schneider-Kamp and Jacob Nielsen and Gianluca Barmina and Kenneth Enevoldsen and Lukas Galke Poech},
4 year={2026},
5 eprint={2608.13517},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2608.13517},
9}
