DFM-Mimir-FP8

DFM-Mimir FP8

This is the FP8 (float8) quantized version of DFM-Mimir — a 1-billion-parameter HRM-Text language model trained from scratch delivering competitive performance for English and state-of-the-art performance for Danish using only permissible post-training data.

Quantization Details

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "schneiderkamplab/DFM-Mimir-FP8",
    torch_dtype="bfloat16",
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("schneiderkamplab/DFM-Mimir-FP8")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Evaluation

Evaluation settings: Temperature 0, shuffle seed 4242, full test datasets, HuggingFace Transformers inference (not vLLM), HRM builtin chat template. MATH limited to 500 samples due to judge grading bottleneck.

Key finding: FP8 retains near-lossless quality across all 17 benchmarks. Average accuracy drop is just 0.9 points vs paper. On 3 benchmarks (HellaSwag, GSM8K, MATH) FP8 exceeds the paper baseline.

English benchmark results (Best scores in bold.)

Math & Code benchmark results (Best scores in bold.)

Danish benchmark results (Best scores in bold.)

Model details

Technical Report

Training was performed using a fork of HRM-Text. Further details are provided in our technical report here.

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.

See the full license text in LICENSE.

Project partners & funding

The development of Mimir v1 was performed in close collaboration between University of Southern Denmark, Aarhus University, University of Copenhagen and the Alexandra Institute, as part of Danish Foundation Models.

Funding was provided by the Ministry of Science, Higher Education and Digital Affairs.

How to cite

@misc{mimir-v1,
  title  = {DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data},
  author = {Schneider-Kamp, Peter and Nielsen, Jacob and Barmina, Gicanluca and Enevoldsen, Kenneth and Poech, Lukas Galke},
  year   = {2026},
  url    = {https://huggingface.co/danish-foundation-models/HRM-Mimir-v1}
}