Instructions to use danish-foundation-models/munin-2.0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use danish-foundation-models/munin-2.0 with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="danish-foundation-models/munin-2.0", trust_remote_code=True) messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("danish-foundation-models/munin-2.0", trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained("danish-foundation-models/munin-2.0", trust_remote_code=True, 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=256) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use danish-foundation-models/munin-2.0 with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "danish-foundation-models/munin-2.0" # 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/munin-2.0", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/danish-foundation-models/munin-2.0
- SGLang
How to use danish-foundation-models/munin-2.0 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/munin-2.0" \ --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/munin-2.0", "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/munin-2.0" \ --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/munin-2.0", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use danish-foundation-models/munin-2.0 with Docker Model Runner:
docker model run hf.co/danish-foundation-models/munin-2.0
Munin 2.0
Munin 2.0 is a 120B-parameter (12B active) open-weights model from Danish Foundation Models for Danish and English chat, reasoning and agentic software engineering. We post-trained it from NVIDIA's pretrained Nemotron 3 Super Base: supervised fine-tuning first, then reinforcement learning on software-engineering and terminal tasks. Full results and background are in the release note.
| Parameters | 120B total, 12B active |
| Architecture | Nemotron 3 Super (unchanged): hybrid Mamba-2 + attention, latent mixture-of-experts, multi-token prediction |
| Languages | Danish and English |
| Context length | 131,072 tokens (128k), as in training |
| Reasoning | On by default; turn it off with enable_thinking=False |
| License | NVIDIA Nemotron Open Model License |
| Model | What it is |
|---|---|
| Munin 2.0 (this repo) | SFT + RL, BF16 weights. |
| Munin 2.0 NVFP4 | Recommended on Blackwell GPUs, which run FP4 natively. |
| Munin 2.0 SFT | The SFT checkpoint Munin 2.0 was trained from. For research and further fine-tuning. |
Highlights
- Agentic coding, the focus of our RL training:
- SWE-bench Verified: 66.2%, against 58.1% for the SFT checkpoint and 49.3% for Nemotron 3 Super in the same harness.
- Terminal-Bench: 13–16 points above the SFT checkpoint.
- Danish: 71.1 on 17 EuroEval datasets, against 68.7 for Nemotron 3 Super. We left out every dataset whose test items appear in our training data; see the release note.
- Danish writing: third of 13 models in a blind comparison, level with or just behind Nemotron 3 Super depending on the judge.
- Reasoning: GPQA-Diamond and AIME improved over the SFT checkpoint, and are level with Nemotron 3 Super within noise.
Quickstart
Tested with vLLM 0.28.0. --quantization fp8_per_block quantizes the BF16 weights to FP8 at load time, which
is how we ran all our evaluations. Drop the flag to serve in BF16. On Blackwell GPUs, use the NVFP4 repo.
vllm serve danish-foundation-models/munin-2.0 \
--served-model-name munin-2.0 \
--trust-remote-code \
--quantization fp8_per_block \
--tensor-parallel-size 8 --enable-expert-parallel \
--max-model-len 131072 \
--kv-cache-dtype fp8 \
--mamba-ssm-cache-dtype float32 \
--async-scheduling --enable-chunked-prefill \
--reasoning-parser nemotron_v3 \
--enable-auto-tool-choice --tool-call-parser qwen3_coder \
--speculative-config '{"method": "mtp", "num_speculative_tokens": 3}'
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
response = client.chat.completions.create(
model="munin-2.0",
messages=[{"role": "user", "content": "Forklar kort forskellen på en kommune og en region."}],
temperature=0.8,
top_p=0.95,
max_tokens=16384,
)
print(response.choices[0].message.content)
- Reasoning off: add
extra_body={"chat_template_kwargs": {"enable_thinking": False}}. - Coding agents: also pass
"force_nonempty_content": True, so a response that ends inside its reasoning still returns content.
Usage notes
- Sampling: use
temperature=0.8andtop_p=0.95. These are thegeneration_config.jsondefaults, and we used them for training and evaluation. At 1.0, which NVIDIA recommends for Nemotron 3 Super, we saw more turns that fall into repetition. - Reasoning across turns: our chat template is NVIDIA's except that
truncate_history_thinkingdefaults toFalse, so earlier turns keep their reasoning, as in RL training. Set it toTrueto save context in long chats; we have not evaluated that setting. - Tool calling: use the
qwen3_coderparser (XML tool calls). During RL we used the opencode and pi coding agents and a plain bash harness. - Output budget: during RL each call could produce up to 16,384 tokens and each episode up to 300,000, with context compaction. Give agents a similar budget.
Training
- Supervised fine-tuning:
- One epoch on 16 LUMI nodes: 6.8M conversations and 11B tokens, of which the 3.4B in assistant turns were trained on.
- About 94% of the trained tokens come from public datasets and 6% from synthetic data we generated.
- About 8% of the trained tokens are Danish and another 9% mix Danish and English.
- It includes the training splits of 24 Danish EuroEval tasks.
- A summary of the training data, following the EU AI Act template, will be published later.
- Reinforcement learning:
- Trained with asynchronous GRPO in our fork of prime-rl, on Gefion (DGX H100).
- About 119,000 episodes from 22 task sources: software engineering (47%), terminal tasks (25%), tool use, SQL, instruction following and maths.
- Every coding and terminal episode ran in its own sandbox, on a sandbox service we built for this project.
- All rewards come from programs: there is no human feedback, no judge model and no Danish data.
- The released checkpoint was chosen partly on the agentic benchmarks below, so its scores on them are somewhat optimistic.
- A summary of the RL training content will be published later. The run history, with every change dated by
step, is on the prime-rl
munin-2.0branches.
- Compute: about 1.5×1021 FLOP in total (SFT 8.7×1020, RL 6×1020).
Evaluation
All three models were served the same way: vLLM, FP8, temperature 0.8, reasoning on. Agentic scores pool the bash, opencode and pi harnesses. ± values are 95% bootstrap intervals over items. Bold marks a difference whose interval excludes zero.
| Benchmark | Munin 2.0 | Munin 2.0 SFT | Nemotron 3 Super | Munin 2.0 − Nemotron (95% CI) |
|---|---|---|---|---|
| EuroEval Danish, 17 datasets ¹ | 71.1 | 66.6 | 68.7 | +2.4 |
| SWE-bench Verified (128 tasks) ² | 66.2 ± 7.2 | 58.1 ± 6.8 | 49.3 ± 7.2 | +16.9 [+11.9, +21.9] |
| Terminal-Bench 2.0 ² | 35.8 ± 8.0 | 23.0 ± 7.9 | 36.0 ± 8.6 | −0.2 [−6.3, +5.9] |
| Terminal-Bench 2.1 ² | 38.1 ± 8.4 | 24.8 ± 7.3 | 34.0 ± 8.4 | +4.1 [−1.1, +9.9] |
| Terminal-Bench Lite ² | 64.2 ± 8.6 | 48.1 ± 8.5 | 56.1 ± 9.3 | +8.1 [+0.5, +15.5] |
| AIME 2025 | 79.2 ± 11.7 | 74.2 ± 11.7 | 76.7 ± 14.2 | +2.5 [−5.8, +12.5] |
| AIME 2026 | 80.0 ± 11.7 | 71.7 ± 11.2 | 76.7 ± 13.8 | +3.3 [−4.2, +10.8] |
| GPQA-Diamond | 72.5 ± 5.3 | 67.2 ± 5.6 | 70.2 ± 5.8 | +2.3 [−2.5, +7.1] |
| Humanity's Last Exam ³ | 10.6 ± 1.3 | 9.2 ± 1.2 | 11.3 ± 1.3 | −0.7 [−2.2, +0.7] |
| IFBench | 49.7 ± 5.7 | 47.0 ± 5.7 | 74.0 ± 5.0 | −24.3 [−30.3, −18.0] |
- EuroEval 18.4, test split. We excluded the 10 of 27 datasets whose test items appear in our training data, and we trained the SFT model on the training splits of most of the remaining ones. The scan covers our own SFT and RL data, not NVIDIA's training data. All of this is explained in the release note, along with a comparison with DFM Mimir.
- Nemotron 3 Super makes few tool calls in our harnesses, so it scores lower here than in NVIDIA's own evaluation. NVIDIA reports 60.5% on SWE-bench Verified (OpenHands) and 31.0% on Terminal-Bench 2.0.
- Judged by Nemotron 3 Super, with answers capped at 32k tokens.
The contamination checks for these benchmarks and a Danish creative-writing comparison are in the release note.
Safety
| Benchmark | Munin 2.0 | Munin 2.0 SFT | Nemotron 3 Super | Better |
|---|---|---|---|---|
| XSTest, correct refusal or compliance | 93.1 | 94.9 | 90.4 | ↑ |
| SABER agent tasks with a harmful action, Danish / English request ⁴ | 42.6 / 41.6% | 44.4 / 51.8% | 49.3 / 51.1% | ↓ |
| Agentic Misalignment, blackmail / leaking | 20 / 79% | 5 / 56% | 15 / 95% | ↓ |
| WMDP-cyber, hazardous cyber knowledge | 67.8 | 60.6 | 62.2 | dual use |
- SABER's user requests are written in Chinese, and we translated them to Danish and English. The rest of each task is in English. gpt-oss-120b judges harm, so the numbers aren't comparable with the SABER paper.
- Better: Munin 2.0 takes harmful actions in agent tasks less often than Nemotron 3 Super.
- Worse: it acts against its operator more often than the SFT checkpoint when it is threatened.
- Dual use: its cyber-security knowledge went up during RL.
We did no dedicated safety training. Add your own safeguards, especially for autonomous agents with real permissions.
Intended use and limitations
- Intended for:
- Danish and English assistants;
- agentic software engineering and terminal work in sandboxes;
- research.
- Not for:
- high-stakes decisions without human oversight;
- running generated code or agent actions outside a sandbox.
- Known limitations:
- Mistakes: it can hallucinate and write wrong code, including wrong fixes that it presents confidently.
- Instruction following: weaker than Nemotron 3 Super's on strict constraints (IFBench).
- Safety: when it is threatened, it blackmails and leaks data more often than the SFT checkpoint (see Safety).
- Agent runs: long runs can produce runaway, repetitive turns, so set per-call token limits.
- Context: most SFT tokens come from short sequences, and we have not tested quality near the end of the 128k window.
- Languages: we have not tested languages other than Danish and English.
- Bias and filtering: we have not measured bias for Danish, and we ran no personal-data or CSAM filtering of our own on the SFT data.
EU AI Act and copyright
The Commission's guidelines treat whoever modifies a model as its provider when the modification uses more than one third of the base model's training compute. Our post-training used about 1.5×1021 FLOP, roughly 0.25% of that threshold, so we do not consider ourselves the provider of a new general-purpose AI model. We will nevertheless publish training-content summaries for the SFT and RL stages. Rights holders can contact the Alexandra Institute, which hosts this work within Danish Foundation Models, at alexandra@alexandra.dk.
License
NVIDIA Nemotron Open Model License; see LICENSE. Licensed by NVIDIA Corporation under the NVIDIA Nemotron Model License.
Project partners & funding
Munin 2.0 was developed as part of Danish Foundation Models, a collaboration between the University of Southern Denmark, Aarhus University, the University of Copenhagen and the Alexandra Institute. The project is funded by the Ministry of Science, Higher Education and Digital Affairs.
The supervised fine-tuning ran on LUMI under DeiC grant DeiC-SDU-N5-2026185. The reinforcement learning ran on Gefion.
How to cite
@misc{dfm2026munin2,
title = {Munin 2.0},
author = {{Danish Foundation Models}},
year = {2026},
howpublished = {\url{https://huggingface.co/danish-foundation-models/munin-2.0}}
}
- Downloads last month
- 445
