borekboissy/Millesime-2026-4b
Millesime-2026-4b by borekboissy is a 4 billion parameter language model, fine-tuned from Qwen3-4B-Instruct-2507, specifically optimized for the French language. It achieves the best FR-MT-Bench score within its evaluation panel, outperforming larger 8B and 9B models, and leads 4B models on aggregated French benchmarks. This model is designed for high performance in French language tasks, particularly excelling in instruction following.
Loading preview...
Millesime 2026 4B: A French-Specialized Small Language Model
Millesime 2026 4B, developed by borekboissy, is a 4 billion parameter model built upon Qwen/Qwen3-4B-Instruct-2507. It is meticulously fine-tuned for excellence in the French language, aiming to surpass models of equivalent size on key francophone benchmarks. A core philosophy of the Millesime project is to focus on Small Language Models (SLMs) that are efficient and can be run locally, while ensuring a clean data chain by only using fine-tuning authorized datasets.
Key Capabilities & Features
- Superior French Performance: Achieves the highest FR-MT-Bench score among its evaluation panel, even outperforming 8B and 9B models, and leads 4B models on aggregated French academic benchmarks.
- Extensive Context Window: Features a large context window of 262,144 tokens.
- Robust Training Pipeline: Utilizes a three-phase pipeline: Supervised Fine-Tuning (SFT) for French general knowledge, DPO for alignment with real human preferences from the Compar:IA dataset, and TIES model fusion.
- Carbon-Efficient: The entire training process consumed only 3.47 kWh, resulting in 1.28 kg CO₂eq emissions.
- Availability: Available in
safetensorsformat, on Ollama, and in various GGUF quantizations for local deployment.
When to Use This Model
- French Language Applications: Ideal for any application requiring high-quality French text generation, comprehension, or instruction following.
- Resource-Constrained Environments: Its 4B parameter size makes it suitable for deployment on devices with limited computational resources.
- Instruction Following: Excels in tasks requiring precise adherence to instructions, as demonstrated by its leading IFEval-fr score.
- Ethical Data Sourcing: For projects prioritizing models trained exclusively on data with explicit fine-tuning authorization.
Limitations
- May be verbose on simple questions due to SFT dataset bias.
- Not designed for explicit chain-of-thought reasoning (non-thinking mode).
- Limited tool-calling capabilities.
- Performance on pure encyclopedic knowledge benchmarks (GPQA-fr, Global-MMLU-fr) is lower than larger models.