Base model: nvidia/NemotronH-30B-A3B (Mamba/MoE hybrid, 31.6B total / 3.58B active, 6-of-128 routing)
Languages: Greek and English (the reasoning trace follows the question's language)
Decoding:thinking: the reasoning trace is carried in <think> blocks and follows the question's language; direct mode is clean (zero unrequested traces across the study's <think>-carrier direct-mode generations, n=9,751)
Serve thinking, and mind one environment note. Under the library versions the study used
(transformers 5.12.1), the shipped remote modeling code trains correctly but fails at
generation (prepare_inputs_for_generation indexes a cache_position that generate()
passes as None); the study evaluated with trust_remote_code=False. mamba-ssm is
optional: installing it changed step time by 0%.
Intended use
Greek and English reasoning where answer-channel integrity matters most: 0.0% leak, and negligible fallback on math and commonsense (1.7 / 1.8%)
Deployments that also need the base's general Greek ability: the fine-tune gains +3.8 Greek macro on the Titan-1 suite (largely catch-up against a low base)
It does not obey an instruction to reason in English on a Greek question (0.0% override; the reverse direction works at 87.5%). Do not deploy it where an operator must force English traces. Not evaluated for safety-critical, legal, or medical use.
Fine-tuned from NemotronH-30B-A3B; shipped merged to dense weights.
Evaluation
Arrows mark the desirable direction for every metric: ↑ higher is better, ↓ lower is better.
Answer-format fallback, per domain (math/cs/logic)
8.3 / 6.2 / 6.6%
1.7 / 1.8 / 39.9%
Instructed override, el→EN / en→EL ↑
n/a (already reasons in English)
0.0% / 87.5%
Median trace length, tokens (own tokenizer) ↓
681
638 (parity, 1.07×)
Greek NLU macro, forgetting check (vs own base) ↑
48.6
52.4 (+3.8)
English NLU macro, forgetting check (vs own base) ↑
78.4
76.4 (−2.0)
The claim replicates on a third, non-transformer architecture: the base reasons in Greek on 0% of the benchmark, the fine-tune on 97.4%, at an accuracy cost of −0.6 pp on that instrument, with generation-cap truncation falling 10.8% → 2.6% and measured loops in its traces dropping 15.0 → 0.0 per thousand rows.
Per-benchmark detail
Per-domain, Greek think lane
Domain (Greek think lane)
Base acc ↑
Sophea-Nemo-3-Nano-v1 acc ↑
Base fallback ↓
Sophea-Nemo-3-Nano-v1 fallback ↓
Math (two math slices)
86.6
71.5
8.3%
1.7%
Commonsense (HellaSwag + Winogrande)
49.2
39.7
6.2%
1.8%
Logic (ProofWriter, macro-recall)
26.8
9.8
6.6%
39.9%
Read the logic cell carefully: the 9.8 macro-recall is substantially a format floor, not a capability reading. The release's fallback is negligible on math and commonsense (1.7 / 1.8%) and 39.9% on logic, the one domain where its answer form breaks; logic is also the hardest axis for every arm in the study, base models included. On the override probe this family returns 5–9% of rows too short to score (Qwen: 0%), consistent with a low reasoning ceiling.
Forgetting check: Titan-1 suite (non-reasoning mode)
This suite exists to answer one question: did the Greek fine-tune damage general ability? It did not. Against its own base, across 14 bilingual NLU benchmarks: +3.8 Greek / −2.0 English; the fine-tune gains on the Greek side and the English movement is within the range seed noise alone produces on this class of suite. Read the numbers below as a forgetting check, not a capability leaderboard.
Benchmark (Titan-1 suite)
Base
Sophea-Nemo-3-Nano-v1
ARC-Challenge-el
54.2
59.0
ARC-Easy-el
62.3
70.1
Belebele-el
67.0
78.0
GreekMMLU
58.8
53.3
HellaSwag-el
39.2
47.6
Medical-MCQA-el
27.8
31.2
TruthfulQA-el
31.7
33.2
Winogrande-el
53.3
58.3
MMLU-el
43.2
40.8
Greek macro
48.6
52.4
ARC-Challenge (en)
88.3
82.2
ARC-Easy (en)
96.2
93.4
HellaSwag (en)
67.9
73.1
Winogrande (en)
69.9
72.9
MMLU (en)
69.9
60.5
English macro
78.4
76.4
Register control improves 40 → 44 of 46 and grammaticality roughly doubles, 13 → 27 of 58, driven by the agreement axis (judge-scored counts, directional evidence only).
1from openai import OpenAI
23client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")4resp = client.chat.completions.create(5 model="sophea-nemo-3-nano-v1",6 messages=[{"role":"user","content":"Ένα κατάστημα πουλά στυλό προς 1,50 € το ένα. Πόσο κοστίζουν 12 στυλό;"}],7 temperature=0.7,8)9print(resp.choices[0].message.content)# <think> trace in the question's language, then the answer
Steering the trace language (partial, measured per checkpoint): append "Think step by step IN ENGLISH." to a Greek question, or "Σκέψου βήμα προς βήμα ΣΤΑ ΕΛΛΗΝΙΚΑ." to an English one. On this checkpoint the English suffix is not obeyed (0.0%); the Greek suffix on an English question is obeyed on 87.5% of items.
License
The fine-tuning additions are released under Apache-2.0. The model inherits the base model's license terms (nvidia/NemotronH-30B-A3B); verify the base model's license before use.
Citation
This model is a release artifact of arXiv:2608.17744. If you use it, please cite:
bibtex
1@misc{kirouane2026thinking,
2 title={Thinking in a Low-Resource Language: What SFT Builds, What RL Fixes, What Accuracy Cannot See},
3 author={Ayoub Kirouane and Christos Petrocheilos},
4 year={2026},
5 eprint={2608.17744},
6 archivePrefix={arXiv},
7 primaryClass={cs.CL},
8 url={https://arxiv.org/abs/2608.17744},
9}
