Instructions to use danish-foundation-models/DFM-Mimir with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use danish-foundation-models/DFM-Mimir with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="danish-foundation-models/DFM-Mimir")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("danish-foundation-models/DFM-Mimir")
model = AutoModelForCausalLM.from_pretrained("danish-foundation-models/DFM-Mimir", device_map="auto")
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]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use danish-foundation-models/DFM-Mimir with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "danish-foundation-models/DFM-Mimir"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "danish-foundation-models/DFM-Mimir",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/danish-foundation-models/DFM-Mimir

SGLang

How to use danish-foundation-models/DFM-Mimir with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "danish-foundation-models/DFM-Mimir" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "danish-foundation-models/DFM-Mimir",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "danish-foundation-models/DFM-Mimir" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "danish-foundation-models/DFM-Mimir",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use danish-foundation-models/DFM-Mimir with Docker Model Runner:
```
docker model run hf.co/danish-foundation-models/DFM-Mimir
```

Is this really a 1B model?

by mindplay - opened 20 days ago

Discussion

mindplay

20 days ago

The model gets detected as 2B - and the file is ~3.6GB, while typical 1B models (such as Gemma 3) at the same precision (BF16) are generally in the area of ~2.0GB.

Is it a 1B model or what are you leaving out? 😊

peter-sk

Danish Foundation Models org 18 days ago

Great question. The model has under 1B parameters without the embeddings and the language modelling head. Taking into account the language modelling head, which is approx. 0.4B parameters, we land at 1.3B. The remaining 0.4B are the embeddings.

mindplay

17 days ago

without the embeddings and the language modelling head

when did we start measuring language models without their embeddings or head?

I'm genuinely asking, because I've come across two other projects recently trying to do the same thing, and it's honestly just confusing.

Correct me if I'm wrong, but usually, if you see e.g. "1B at 8BFP", you'd expect to see a ~1GB file, right?

has the whole industry recently decided it's more important to look competitive than it is to provide us with the real dimensions? 😊

it feels like obfuscation, and it makes it difficult to compare... "1B" to begin with can already mean a lot of different things - from your 3.6GB model down to Bonsai/BitNet type models at a few hundred MB.

speaking of, will you be shipping quantized versions, and do you have any idea what sort of performance we can expect there? 16 bits per parameter is a bit high by today's standards for small models? 😊

KennethEnevoldsen

Danish Foundation Models org 17 days ago

I am sorry if you believe we have been misleading. It is not completely uncommon, many models vary in their reported size by quite a bit and we are by no means the first to focus on the active parameters and discount the embedding layer.

E.g. over at MTEB we report active parameters as it closer reflects the runtime and ecologits use active parameter to estimate energy usage.

Huggingface luckily allow you to inspect the dimensions on each layer, so it the numbers should be there for you to inspect.

I don’t know of any plans to make a low precision version of the model.

pathosethoslogos

14 days ago

•

edited 14 days ago

redacted

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment