Base model: nvidia/Nemotron-3.5-Lightning-30B-A3B (next-gen 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. Leave thinking enabled so the trace is produced; it follows the
question's language. Family environment note: the study's NemotronH sibling required
trust_remote_code=False at generation time under transformers 5.12.1; verify generation on
your stack before deploying.
Intended use
Greek and English reasoning on the strongest Nemotron-line base measured in the study (Greek macro 57.5 vs Nano's 48.6), with retention gains on Greek (+1.7) and a 1.1-point English cost
Deployments wanting the hybrid family with the mildest logic fallback (5.3%)
It does not obey an instruction to reason in English on a Greek question (0.0% override; the reverse direction works at 92.7%). Do not deploy it where an operator must force English traces. Not evaluated for safety-critical, legal, or medical use.
Fine-tuned from Nemotron-3.5-Lightning-30B-A3B, released as a third matched checkpoint of the language-matched recipe, not a full replication of the Qwen and Gpt-OSS studies.
Evaluation
Arrows mark the desirable direction for every metric: ↑ higher is better, ↓ lower is better.
Answer-format fallback, per domain (math/cs/logic)
1.4 / 1.9 / 4.4%
16.1 / 14.7 / 5.3%
Instructed override, el→EN / en→EL ↑
n/a (already reasons in English)
0.0% / 92.7%
Greek NLU macro, forgetting check (vs own base) ↑
57.5
59.2 (+1.7)
English NLU macro, forgetting check (vs own base) ↑
82.0
80.9 (−1.1)
Fidelity measures 98.06% on the identical probe (printed as 98.1 in the paper's family table), with zero in-question switches and 100% English-control compliance.
Per-benchmark detail
Per-domain, Greek think lane
Domain (Greek think lane)
Base acc ↑
Sophea-Nemo-3.5-Lightning-v1 acc ↑
Base fallback ↓
Sophea-Nemo-3.5-Lightning-v1 fallback ↓
Math (two math slices)
76.9
75.6
1.4%
16.1%
Commonsense (HellaSwag + Winogrande)
46.3
40.9
1.9%
14.7%
Logic (ProofWriter, macro-recall)
25.0
23.8
4.4%
5.3%
Logic is a capability floor shared with its own base (23.8 vs 25.0 macro-recall; logic is the hardest axis for every arm in the study), not a fine-tuning regression: the release's logic fallback is a mild 5.3%, the lowest of the five releases. Its fallback cost sits instead on math and commonsense (16.1 / 14.7%). Like its Nano sibling, this family returns 5–9% of override-probe rows too short to score.
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: +1.7 Greek / −1.1 English, both 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-3.5-Lightning-v1
ARC-Challenge-el
68.7
72.8
ARC-Easy-el
75.8
81.4
Belebele-el
81.3
83.3
GreekMMLU
64.2
62.2
HellaSwag-el
44.5
51.2
Medical-MCQA-el
41.7
38.2
TruthfulQA-el
34.2
33.7
Winogrande-el
55.8
60.1
MMLU-el
51.3
49.6
Greek macro
57.5
59.2
ARC-Challenge (en)
92.0
89.1
ARC-Easy (en)
97.8
96.4
HellaSwag (en)
72.7
75.9
Winogrande (en)
72.0
73.3
MMLU (en)
75.8
69.7
English macro
82.0
80.9
Register control improves 37 → 43 of 46 and grammaticality roughly doubles, 12 → 29 of 58 (judge-scored counts, directional evidence only). In direct (non-reasoning) mode on the same suite the release reads +1.7 against its own base.
1from openai import OpenAI
23client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")4resp = client.chat.completions.create(5 model="sophea-nemo-3.5-lightning-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 92.7% of items.
Speculative decoding (MTP)
This model ships the base's multi-token-prediction head: 270 mtp.* tensors (~2.7 GB, bf16)
in model-mtp.safetensors, the draft stack that config.json declares through
num_nextn_predict_layers: 1 (an attention block plus a MoE block). The head is the base
model's own; speculative decoding verifies every drafted token against the main model, so it
changes throughput only, never the output distribution.
Per vLLM's guidance for this family, MTP suits low-latency serving; for maximum batch
throughput, run without speculative decoding.
License
The fine-tuning additions are released under Apache-2.0. The model inherits the base model's license terms (nvidia/Nemotron-3.5-Lightning-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}
