Luth-2-0.8B

Luth-2-0.8B is a 750M-parameter (text only) non-reasoning model, setting a new state of the art in French for its size across math, code, instruction following, general knowledge and tool calling. It is trained on a 3B-token French SFT mixture followed by multi-domain on-policy distillation (MOPD). The model outperforms every other model in its size class on our selected French benchmarks and stays competitive with models 2 to 3 times larger. It is small enough for efficient local and on-device deployment.

• 📄 Blog: Luth-2: Pushing the French Capabilities of SLMs with MOPD
• 🤗 Models: Luth-2-0.8B · Luth-2-2B
• 📊 Datasets: SFT · RL
• 💻 Code: GitHub
• 🏆 Leaderboard: French LLM Leaderboard

Luth-2-0.8B inherits the VLM architecture of Qwen3.5-0.8B but was not trained on vision data. We do not recommend using it for vision tasks.

Model variants

Model	Description
Luth-2-0.8B	Original checkpoint in native format. Best for fine-tuning or inference with Transformers, vLLM and SGLang.
Luth-2-0.8B-GGUF	Quantized format for llama.cpp and compatible tools. Optimized for CPU inference and reduced memory usage.

Training

Luth-2-0.8B is post-trained from Qwen3.5-0.8B in two stages:

1. Supervised fine-tuning on Luth-2-Post-Training-SFT, a 3B-token French mixture spanning math (37.2%), knowledge (27.9%), code (22.2%), instruction following (6.5%) and tool calling (6.3%). Prompts were translated from English SFT datasets and answers regenerated with strong open-source teachers.
2. Multi-domain on-policy distillation (MOPD). Three specialists (math, code, instruction following) are trained separately with GRPO on Luth-2-Post-Training-RL, then distilled back into the SFT student.

Inference

Luth-2-0.8B is supported by Transformers, vLLM, SGLang and more.

Quick start with Transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

model_id = "kurakurai/Luth-2-0.8B"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype="bfloat16",
    # attn_implementation="flash_attention_2"  # uncomment on compatible GPU
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

prompt = "Quelle est la capitale de la France?"
input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
)["input_ids"].to(model.device)

output = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.8,
    top_p=0.95,
    top_k=20,
    max_new_tokens=512,
    streamer=streamer,
)

Evaluation

Evaluations can be reproduced using our GitHub repository. The benchmarks are French subsets or verified translations, scored with temperature=0.6, top_p=0.95, top_k=20, thinking disabled, averaged over 10 runs.

French Benchmarks	Luth-2-0.8B	Luth-0.6B-Instruct	Qwen3.5-0.8B
MGSM-rev2	72.92	58.52	35.20
AIME 24	5.67	2.00	1.00
AIME 25	8.67	1.33	0.33
Math-500	57.60	44.74	27.46
Global-MMLU-Lite	53.30	40.20	44.00
MMLU-ProX-Lite	38.93	25.40	27.60
GPQA-Diamond	26.87	25.60	23.80
IFEval	71.23	51.23	44.47
Multi-IF	61.52	33.77	32.72
HumanEval+	46.81	30.25	10.87
MBPP+	42.33	34.74	18.20
BFCL v2	64.02	61.72	51.49

See the French LLM Leaderboard for comparisons across models.

Contact

Questions or feedback? Reach us on LinkedIn: Maxence Lasbordes and Guillaume Pradel.

Citation

@misc{luth2,
  title  = {Luth-2: Pushing the French Capabilities of SLMs with MOPD},
  author = {Maxence Lasbordes and Guillaume Pradel},
  year   = {2026},
  url    = {https://huggingface.co/blog/MaxLSB/luth-2}
}