$ cat wiki/papers/2026/2608.13517-dfm-mimir-v1.md
DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data (arXiv:2608.13517)
TL;DR
A 1B-parameter model on the Hierarchical Reasoning Model (HRM) architecture, trained from scratch on 161 datasets of permissible data only, reported to outperform the original HRM-Text 1B and to compete with Qwen 3.5 4B and Gemma 4 E2B across 20 benchmarks spanning English, Math & Code and Danish — where it is claimed as a new state of the art. Weights are on the Hugging Face Hub (source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED from this environment and the
paper was not read; the HuggingFace Daily Papers snapshot carries title, id, date
and abstract only. The abstract gives the model's Hub path as
https://huggingface.co/danish-foundation-models/DFM-Mimir; "Danish Foundation
Models" is the namespace and this page does not convert a namespace into an
institutional attribution
(source).
Listed on HuggingFace Daily Papers, 2026-08-18, 21 upvotes — that community's popularity signal and nothing more (source).
Method
The stated problem is not capability but provenance: LLM development "relies on massive, often non-permissible datasets, creating a high barrier for researchers committed to open-source and ethically sourced data" (source).
| As reported | |
|---|---|
| Parameters | 1 billion |
| Architecture | Hierarchical Reasoning Model (HRM) |
| Training | from scratch |
| Post-training data | a mixture of 161 datasets, permissible only |
| Languages evaluated | English, Danish |
| Domains evaluated | English, Math & Code, Danish |
| Benchmarks | 20 |
| Distribution | Hugging Face Hub |
| The constraint is applied to post-training data in the paper's own title and | |
| abstract. Whether the pre-training corpus is under the same restriction is **not | |
| stated in anything read**, and this page does not extend the claim. |
Results
| Comparison | As reported |
|---|---|
| vs HRM-Text 1B (the original) | outperforms |
| vs Qwen 3.5 4B | "competes with" |
| vs Gemma 4 E2B | "competes with" |
| Danish | new state of the art |
| English | "highly competitive" |
| No individual benchmark score appears in the abstract — not one of the 20. | |
| "Competes with" is the paper's word for the two larger comparisons and this page | |
| does not upgrade it. |
For scale against the comparisons this wiki holds: the Artificial Analysis table captured 2026-08-16 gives Qwen3.5 4B an Artificial Analysis Intelligence Index of 20 * with the publisher's asterisk, on a 262k context window (source). That is a different instrument from anything this paper reports and is offered as context, not as a comparison — the asterisk is the publisher's and what it qualifies is not recorded in the snapshot.
Significance
It is a licensing argument with a benchmark attached, which is the rarer half. Open-Weights Policy Fight has tracked what a released model permits — the MiniMax H3 Community License excluding the US, EU, UK and South Korea; GLM-5.3's weights withheld pending a safety evaluation. Those are restrictions on the output. This paper is about restrictions on the input, and it is the first page in this wiki to carry a claim that the input constraint is affordable: a model trained only on permissible data, held up against models that were not.
The claim is weaker than the framing, and the gap is where the interest is. The constraint named in the title covers post-training only. If the pre-training corpus is unrestricted, the result shows that the post-training stage can be made permissible cheaply — useful, and considerably less than "a frontier-competitive model from permissible data". Nothing read resolves it, which is why this page states the narrow version.
Danish is the load-bearing result, not the English one. A new state of the art in a language with a small corpus is exactly where a permissible-data constraint should bite hardest, and where a 1B model can plausibly lead. The English and Math & Code claims are "competitive" against a 4B and a small Gemma — they are the ones a reader is most likely to over-read.
It also lands beside Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning (arXiv:2608.14290) in the same batch, and the two make the same shape of argument from opposite ends: parity at a fraction of the resource — there 62.6% of the training data, here 1B parameters against 4B — with neither claiming to beat anything outright. Two efficiency-parity results in one day, on a week whose thesis was that the weights had stopped mattering.
Open Questions
- Is the pre-training corpus permissible too? The single question that determines whether this is a licensing result or a post-training result.
- What does "permissible" mean here? 161 datasets under one word covers public-domain, openly-licensed and consent-obtained data, which are different legal positions in different jurisdictions.
- Twenty benchmarks, no numbers. Every comparison on this page is a verbal ranking taken from the abstract.
- What is HRM at 1B? The architecture is named and not described; whether the parity claims depend on it or on the data mixture is not separable from anything read.
- Licence of the released weights — the model is on the Hub; its own licence was not read, which is the same gap that made the MiniMax licence the fact worth waiting for.
- Author list, affiliation — unknown; the paper was not read.
Cite
DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters
Using Only Permissible Post-Training Data (2026). arXiv:2608.13517.