Skip to content
Language models

DFM-Mimir-v1.8-Mini

a compact multilingual model on the HRM-Text architecture

2 min readupdated

DFM-Mimir-v1.8-Mini is a compact 703.6M-parameter text generation model from Danish Foundation Models. Unlike standard transformers, it uses the HRM-Text architecture, combining hierarchical recurrent computation with a PrefixLM training objective (a mode that blends bidirectional prefix context with autoregressive generation). Trained on the DFM15 (MTC v1.8) data mix—including DBC and Lex.dk materials—and evaluated across 50 languages, it's a solid fit for researchers exploring non-standard recurrent architectures and developers looking for a lightweight multilingual model under a permissive Apache-2.0 license.

Key features

  • HRM-Text architecture: configured with H=2 and L=3 parameters, the full 262,144-token Gemma tokenizer, and thinking mode disabled.
  • Context and weights: features a base context length of 4,096 tokens (extended to 8,192 via parameter overrides during evaluation), with weights published in BF16 (final epoch 3 EMA checkpoint at step 1,538,766).
  • 50-language evaluation: benchmarked across 396 metrics in 50 languages, with included zero-shot and few-shot EuroEval results.
  • Custom runtime required: needs an environment with HRM and PrefixLM support (tested with modified builds of Transformers 5.17.0 and vLLM 0.31.0).

What you can do

  • Explore and test HRM-Text: spin up a local OpenAI-compatible API server via vLLM to experiment with hierarchical recurrent generation.
  • Handle multilingual text tasks: generate and process text across European and other languages from its 50-language evaluation suite, including Danish and English.
  • Run comparative benchmarks: reproduce EuroEval protocols or cross-check model behavior against the included evaluation_results.json reports.

Example

python -m vllm.entrypoints.openai.api_server \
  --model ./DFM-Mimir-v1.8-Mini \
  --tokenizer ./DFM-Mimir-v1.8-Mini \
  --served-model-name danish-foundation-models/DFM-Mimir-v1.8-Mini \
  --chat-template ./DFM-Mimir-v1.8-Mini/chat_template.jinja \
  --host 127.0.0.1 --port 8200 \
  --attention-backend FLASH_ATTN --max-model-len 8192
Read moreDetails
Average score across 396 metrics in 50 languages is 66.0075 (66.3236 excluding English and Danish), outperforming the step 1,500,000 checkpoint on average score in 33 out of 50 languages.
Standard Transformers and vLLM builds without HRM support will not work, and the repository does not provide remote-code auto-loading.
Keep fix_mistral_regex=false at launch to avoid breaking tokenization.
Model weights are licensed under Apache-2.0 (separate Gemma tokenizer terms apply).
Discuss Telegram 0 100

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 1.3 GB · ready files2.1 GB
  • FP8 / INT8recommendedweights 0.7 GB1.4 GB
  • 4-bitweights 0.4 GB1.2 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Language models: with a 8192-token context.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

A ready script for the machine we advise above: paste it and the model downloads and starts.

  1. 1. Rent A30

    $0.26 an hour with the coupon. Pick the Ubuntu image when you create the machine.

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # danish-foundation-models/DFM-Mimir-v1.8-Mini — A30, FP8 / INT8
    # NVIDIA driver: CUDA 13 needs 580+. If older or missing, the current nvidia-open is installed (keep yours: KEEP_DRIVER=1)
    DRV=$(nvidia-smi --query-gpu=driver_version --format=csv,noheader 2>/dev/null | head -1 | cut -d. -f1)
    case "$DRV" in ""|*[!0-9]*) DRV=0 ;; esac
    if [ "$DRV" -lt 580 ] && [ -z "${KEEP_DRIVER:-}" ]; then
      . /etc/os-release && \
      curl -fsSLO "https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/x86_64/cuda-keyring_1.1-1_all.deb" && \
      sudo dpkg -i cuda-keyring_1.1-1_all.deb && rm -f cuda-keyring_1.1-1_all.deb && \
      sudo apt-get update && \
      sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-open || \
      { echo "Could not install the NVIDIA driver: see the error above. To keep the current one, run with KEEP_DRIVER=1."; exit 1; }
      echo "NVIDIA driver updated. Reboot the machine (sudo reboot) and run this same script again."
      exit 0
    fi
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv venv ~/vllm --python 3.12 && source ~/vllm/bin/activate
    uv pip install vllm
    vllm serve danish-foundation-models/DFM-Mimir-v1.8-Mini --max-model-len 8192 --quantization fp8
    Add more LoRAs to DFM-Mimir-v1.8-Mini
  3. 3. Use it

    The server listens on port 8000 and speaks the OpenAI API: http://localhost:8000/v1 on the machine. From your computer: ssh -L 8000:localhost:8000 user@machine-IP.

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

New posts about versions of DFM-Mimir-v1.8-Mini (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

danish-foundation-models/DFM-Mimir-v1.8-Mini
Author
danish-foundation-models
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

1000

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments