DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data
A 1B-parameter HRM model is claiming competitive English results and state-of-the-art Danish performance using only permissible post-training data.
Mimir v1 was trained from scratch on a mix of 161 datasets, according to the arXiv abstract. The authors say it beats the original HRM-Text 1B and competes with larger models including Qwen 3.5 4B and Gemma 4 E2B across 20 English, math, code, and Danish benchmarks. The latest version adds memorization audits. HF Daily Papers' note
Mimir v1 was trained from scratch on a mix of 161 datasets, according to the arXiv abstract. The authors say it beats the original HRM-Text 1B and competes with larger models including Qwen 3.5 4B and Gemma 4 E2B across 20 English, math, code, and Danish benchmarks. The latest version adds memorization audits. HF Daily Papers' note
score 6