DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data

The Core Achievement

Mimir v1 challenges the assumption that frontier-level language model performance requires access to massive, legally murky datasets. The authors built a 1-billion-parameter model trained entirely on permissible data—datasets with clear licensing and ethical sourcing—and showed it competes with larger, less-scrupulous peers. On English benchmarks, Mimir v1 matches or beats models like Qwen 3.5 4B and Gemma 2 8B. For Danish, it sets a new state-of-the-art, a significant result for under-resourced languages often neglected in AI development.

How It Works

Mimir v1 uses the Hierarchical Reasoning Model (HRM) architecture, trained from scratch on a curated mixture of 161 datasets. Rather than relying on web-scraped data of unknown provenance, the team assembled datasets with explicit permissions—academic papers, open corpora, and licensed resources. This constraint is tight: permissible data is far smaller and more fragmented than the internet-scale pools most frontier labs use.

The key insight is that diversity and curation can partially compensate for data size. By mixing 161 sources, the model sees varied reasoning patterns, code examples, and language styles. This breadth appears to help the model generalize despite the total volume being constrained by ethical sourcing requirements.

The model was tested across 20 benchmarks spanning English, mathematics, coding, and Danish. Performance was evaluated against established baselines and contemporary peers—the 1B HRM-Text baseline, as well as larger commercial models like Qwen 3.5 4B.

Why This Matters

For research accessibility: Many researchers and smaller organizations cannot or will not use datasets obtained through aggressive web scraping or terms-of-service violations. Mimir v1 demonstrates that competitive performance is achievable under these constraints, lowering barriers to entry.

For language diversity: Most frontier models focus on English. Mimir v1's state-of-the-art Danish performance hints at a path forward for other low-resource languages. If permissible data strategies work for Danish, they may work for other under-served languages with committed communities.

For reproducibility and governance: Models built on openly licensed data are easier to audit, reproduce, and understand. There's no hidden debt from undisclosed training data. This matters for organizations subject to data governance regulations or concerned about intellectual property risks.

The trade-off: The authors' achievement doesn't mean permissible data is free. It requires careful curation, sourcing from academic repositories and licensed corpora, and accepting that you won't have billions of unlabeled web pages. But for many use cases—especially in regulated industries or non-English contexts—that trade-off is worth it.

Mimir v1 is released on Hugging Face, signaling the authors' commitment to transparency and community reuse. The result is incremental, not revolutionary—the model isn't larger or faster than alternatives, and permissible data remains a constraint. But it resets the conversation: frontier performance and ethical sourcing are not mutually exclusive.