⚙️ Open Models

Train from scratch or bring supported Hugging Face checkpoints for language generation, translation, vision, OCR, and speech.

🧱 Simplicity and Modularity

Validated YAML configurations, reusable model components, LoRA fine-tuning, and native translation scorers support experimentation.

💨 Speed and Efficiency

Serve through chat APIs and use optional CUDA kernels, quantization, compilation, and native MTP on compatible models and hardware.