New Small Language Models Released for French Language Tasks

A new series of small language models (SLMs) optimized for the French language, titled Luth-2, has been released. The release includes two non-reasoning models: Luth-2-0.8B and Luth-2-2B. According to reports from developers on Reddit and Hugging Face, these models are designed to be lightweight enough for local, on-device execution.

The developers state that the Luth-2 series utilizes the Qwen3.5 architecture as its backbone, noting that this foundation proved more receptive to post-training processes. The training methodology involved a 3-billion-token Supervised Fine-Tuning (SFT) mixture covering domains such as mathematics, code, tool calling, instruction following, and science. Additionally, the developers employed reinforcement learning through expert specializations and multi-domain on-policy distillation (MOPD) to enhance model capabilities.

Benchmark Performance and Comparative Data

According to unverified claims shared by the developers on Reddit, the Luth-2 models demonstrate performance levels that exceed several larger models on specific French-language benchmarks.

In comparative testing reported by the developers:

  • Luth-2-2B achieved a score of 69.67 on the Multi-IF benchmark, compared to 65.17 for the larger Gemma-4-E2B-it.
  • Luth-2-2B scored 81.52 on the Math-500 benchmark, slightly outperforming the Gemma-4-E2B-it score of 81.24.
  • Luth-2-0.8B recorded a score of 72.92 on MGSM-Rev2, significantly higher than the 55.60 reported for granite-4.0-h-micro.

The developers suggest that these results indicate a performance gap in current multilingual SLMs when operating outside of English, even for high-resource languages like French.

Availability and Community Context

The Luth-2-0.8B and Luth-2-2B models, along with their GGUF versions for quantized use, are currently available on Hugging Face. The developers have also released the associated SFT and RL datasets and provided a GitHub repository containing the code.

The release has prompted discussion regarding the necessity of high-performance, native-language models. Some commentators on Reddit argued that as SLMs become more capable, there is increasing demand for models that can run locally on mobile devices and provide high-quality performance in languages other than English. These discussions highlighted a perceived bias in current AI development, where benchmarks and model optimizations are frequently centered on English-speaking contexts, potentially overlooking the specific requirements of non-English users.