Munin 2.0 SFT

Munin 2.0 SFT is a supervised fine-tune of NVIDIA's pretrained Nemotron 3 Super Base (120B parameters, 12B active), for Danish and English. Danish Foundation Models built it as the starting point for reinforcement learning.

For most uses, choose Munin 2.0 instead. It is this model after RL on agentic coding and terminal tasks, and it scores as well or better on every capability benchmark we ran, including Danish. This checkpoint is released for research and for anyone who wants to run their own RL or fine-tuning on top of a Danish-capable Nemotron 3 Super.

How it was trained

  • Starting point: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-Base-BF16, revision b4cc2020.
  • Data:
    • one epoch over about 100 sources: 6.8M conversations and 11.1B tokens, of which the 3.4B in assistant turns were trained on;
    • trained tokens are roughly 71% English and code, 8% Danish, 9% mixed Danish/English, 12% other languages;
    • about 94% of trained tokens come from public datasets;
    • about 6% is synthetic data we generated;
    • the data includes the training splits of 24 Danish EuroEval tasks.
  • Full data list: a public summary of the training content will be published later.
  • Recipe: prime-rl, 2,638 steps of 32 packed sequences of 131,072 tokens, learning rate 1e-5, on 16 LUMI nodes (AMD MI250X). About 8.7ร—1020 FLOP and 1,870 kWh.
  • Chat template: NVIDIA's, except that truncate_history_thinking defaults to False. Reasoning is on by default; turn it off with enable_thinking=False.

Results

These come from the same pipeline as the Munin 2.0 card: vLLM FP8, temperature 0.8, reasoning on. The ยฑ values are 95% intervals.

Benchmark Munin 2.0 SFT Munin 2.0 Nemotron 3 Super
EuroEval Danish, 17 datasets 66.6 71.1 68.7
SWE-bench Verified (128) 58.1 ยฑ 6.8 66.2 ยฑ 7.2 49.3 ยฑ 7.2
Terminal-Bench 2.0 23.0 ยฑ 7.9 35.8 ยฑ 8.0 36.0 ยฑ 8.6
GPQA-Diamond 67.2 ยฑ 5.6 72.5 ยฑ 5.3 70.2 ยฑ 5.8
AIME 2026 71.7 ยฑ 11.2 80.0 ยฑ 11.7 76.7 ยฑ 13.8
IFBench 47.0 ยฑ 5.7 49.7 ยฑ 5.7 74.0 ยฑ 5.0

See the main card for the full table, contamination checks and safety results. On some safety measures this checkpoint behaves better than Munin 2.0. For example, it blackmails in 5% of Agentic Misalignment runs, against 20% for Munin 2.0.

Usage

Serve it exactly as Munin 2.0. The Munin 2.0 card has the vLLM command (--reasoning-parser nemotron_v3 --tool-call-parser qwen3_coder) and recommended sampling (temperature=0.8, top_p=0.95).

Limitations

The same as Munin 2.0's. It has had no RL, so it is weaker at agentic coding.

License

NVIDIA Nemotron Open Model License. Licensed by NVIDIA Corporation under the NVIDIA Nemotron Model License.

Downloads last month
313
Safetensors
Model size
124B params
Tensor type
F32
ยท
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for danish-foundation-models/munin-2.0-sft

Finetuned
(11)
this model
Finetunes
1 model

Collection including danish-foundation-models/munin-2.0-sft