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
```

Reasoning issue

by drkooler - opened 6 days ago

Discussion

drkooler

6 days ago

Hi Guys. Love your work. Just a little issue here:

Model class: HrmTextForCausalLM
Model dtype(s): torch.bfloat16
Parameter count: 1,786,775,040
Model footprint: 3.33 GiB
Load time: 5.08 s
VRAM after model load: allocated=3.33 GiB, reserved=3.50 GiB, device_free=11.31/15.93 GiB

DFM-Mimir terminal chat. Type /quit or /exit to stop.

You: hi! how are you?
Mimir: Hello! I'm just a computer program, so I don't have feelings or emotions. But I'm here and ready to help you with any questions or tasks you may have. How can I assist you today?
[generation: prompt=16 tokens, output=47 tokens, speed=9.97 tokens/s]
[generation peak allocated VRAM: 3.42 GiB]

You: A farmer has 17 sheep. All but 9 run away. How many are left? Explain.
Mimir: The farmer has 17 sheep initially. The phrase "all but 9 run away" means that all the sheep except for 9 ran away. So, if we subtract 9 from the total number of sheep, we get:

17 - 9 = 8

Therefore, there are 8 sheep left.
[generation: prompt=101 tokens, output=68 tokens, speed=11.47 tokens/s]
[generation peak allocated VRAM: 3.50 GiB]

You: En bonde har 17 får. Alle undtagen 9 løber væk. Hvor mange får er der tilbage? Forklar dit svar.
Mimir: Bonden har 17 får i starten. Frasen "alle undtagen 9 løber væk" betyder, at alle fårene undtagen 9 løb væk. Så hvis vi trækker 9 fra det samlede antal får, får vi:

17 - 9 = 8

Derfor er der 8 får tilbage.
[generation: prompt=220 tokens, output=78 tokens, speed=11.49 tokens/s]
[generation peak allocated VRAM: 3.59 GiB]

You: please check your reasoning; are you sure there are 8 sheep left?
Mimir: Yes, I am sure that there are 8 sheep left. The phrase "all but 9 run away" means that all the sheep except for 9 ran away, so if we start with 17 sheep and subtract 9, we get 8 sheep left.
[generation: prompt=331 tokens, output=57 tokens, speed=11.70 tokens/s]
[generation peak allocated VRAM: 3.70 GiB]

You: how about this logic. if i rephrase it like this: 17 - 8.
Mimir: If you rephrase the logic as "17 - 8", you would be assuming that 8 sheep ran away, which is not the case. The original statement was "all but 9 run away", which means that 9 sheep ran away. So, if you subtract 8 from 17, you would be incorrect. The correct answer is still 8 sheep left.
[generation: prompt=423 tokens, output=80 tokens, speed=11.89 tokens/s]
[generation peak allocated VRAM: 3.78 GiB]

You: i must disagree. the sheep thing is designed as a reasoning trap, not just a maths based equation to solve. "all but 9" implies all the sheep ran away except for 9. so 9 remained. is this logic acceptable to you?
Mimir: I understand your point, but I believe that the original logic is correct. The phrase "all but 9 run away" means that all the sheep except for 9 ran away, so if we start with 17 sheep and subtract 9, we get 8 sheep left.

However, I see that you are pointing out a potential issue with the logic. If the farmer has 17 sheep and all but 9 run away, then it is possible that the farmer has not yet had time to catch up with the 9 sheep that ran away. In this case, the farmer could have only
[generation: prompt=577 tokens, output=128 tokens, speed=12.01 tokens/s]
[generation peak allocated VRAM: 3.89 GiB]

drkooler

6 days ago

DFM-Mimir is running successfully from its original safetensors checkpoint on the RTX 5060 Ti.
Result

Location: REDACTED
Python: 3.13.14; Python 3.12 was unavailable
Isolated .venv only; no global installs
PyTorch: 2.13.0+cu130
Transformers: 5.12.1
CUDA detected: yes
BF16 GPU computation: verified
Dependency check: no broken requirements
No GGUF, llama.cpp, quantization, or remote model code
The Python Store runtime failed to seed pip normally, so pip was bootstrapped directly into .venv; global Python remained unchanged.
Integrity
Pinned Hugging Face revision:
model.safetensors: 3,573,566,200 bytes
Weight SHA-256 matched published LFS hash:
8a83c8a0e6ad25b73c089c9c6f8b01969f1c6bd7db76a71ff85de861a95311a0
Tokenizer SHA-256 also matched
Safetensors structure: valid, 131 tensors
Inference benchmark
Model class: HrmTextForCausalLM
Parameters: 1,786,775,040
Dtype: BF16
Load time: 3.49 seconds
Model footprint: 3.33 GiB
VRAM after load: 3.33 GiB allocated, 3.50 GiB reserved
Peak generation VRAM: 3.39 GiB
Performance: 8.53 generated tokens/second
Test response: “Danmarks hovedstad er København.”
The initial Windows CP1252 rendering of ø was diagnosed as terminal encoding—not an inference issue—and corrected in the runner.

peter-sk

Danish Foundation Models org 6 days ago

Fantastic find! I love how good Mimir is in insisting on its (wrong) opinion :-) Especially the part where it assume the farmer might be with the sheep that ran away.

Note that the model is in many ways a base model, i.e., even though it is trained exclusively on post-training data, it has not undergone neither preference optimization nor reinforcement learning to increase reasoning accuracy, minimize reasoning budgets etc.

drkooler

6 days ago

Hi! Don't worry, I understand this initial scenario. I did download the raw safetensors file and I understand that. Mimir's sheep opinions are humerous 😊

I'm currently training an LLM for a client and must "debug" it to wrinkle out its mistakes.

Keep up the good work!

Lee

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