이미지: METAL LAB 생성
Summary
- kurakurai has released Luth-2-0.8B and Luth-2-2B, small language models specialized for French
- Both models outscored models three times their size on Multi-IF, MGSM-Rev2, and Math-500
- Built on a Qwen3.5 backbone, they combine 3B-token SFT, reinforcement learning, and multi-domain on-policy distillation
- 모델명
- Luth-2-0.8B, Luth-2-2B
- 유형
- non-reasoning 소형언어모델(SLM)
- 백본 모델
- Qwen3.5
- 공개처
- Hugging Face (GGUF 양자화 버전 포함)
- 주요 벤치마크
- Multi-IF 69.67(2B) vs Gemma-4-E2B-it 65.17, MGSM-Rev2 72.92(0.8B) vs granite-4.0-h-micro 55.60
A 2-billion-parameter model beats a 6-billion-plus model
Development team kurakurai has released Luth-2-0.8B and Luth-2-2B, small language models specialized for French. With 800 million and 2 billion parameters respectively, both are compact enough to run on smartphones or laptops. Both models were built as "non-reasoning" models that produce answers directly without going through a separate reasoning process.
The published benchmark scores show a notable gap despite the size difference. Luth-2-2B scored 69.67 on Multi-IF, which measures instruction-following ability, edging out Gemma-4-E2B-it (65.17), a model roughly three times larger. On the math reasoning benchmark MGSM-Rev2, Luth-2-0.8B scored 72.92, far outpacing granite-4.0-h-micro (55.60). On Math-500, Luth-2-2B recorded 81.52, narrowly beating Gemma-4-E2B-it (81.24).
Why did this happen in French
Small language models (SLMs) aim to preserve as much of a large model's performance as possible while shrinking its size. They are a key ingredient for on-device AI that runs on smartphones or personal computers without an internet connection. The problem is that performance often drops sharply in languages other than English — and according to the team, even French, which has relatively abundant training data, was no exception.
Luth-2 changed three things to close this gap. First, the team newly constructed a 3-billion-token supervised fine-tuning (SFT) dataset spanning math, knowledge, code, tool use, instruction following, multi-turn dialogue, and science. On top of this, they added domain-specific reinforcement learning and multi-domain on-policy distillation (MOPD) — a training method that transfers a larger model's latest judgments to a smaller model in real time. The backbone was also switched from the previous generation's base model to Qwen3.5, which the team said responded much better to post-training.
The trend of building derivative models tailored to specific languages or domains on top of the Qwen family as a backbone has been notable in recent weeks. Qwen's team releasing Qwen3-ASR-0.6B, a lightweight speech recognition model supporting 30 languages and 22 Chinese dialects, in early August fits the same pattern.
Qwen releases lightweight speech recognition model Qwen3-ASR-0.6B
Qwen releases lightweight speech recognition model Qwen3-ASR-0.6B
So what changes now
Luth-2 was released not just as model weights but together with the SFT and RL datasets used in training and the training code itself. A leaderboard dedicated to comparing French-language performance has also been set up, opening the door for developers to reproduce the results directly or apply the same methodology to other languages. In particular, the fact that an 800-million-parameter model beat a model nearly five times its size on a math benchmark suggests that fine-grained, language-specific tuning can narrow performance gaps without simply scaling up model size.
