Steerling-8B, a new 8-billion-parameter language model, achieves unprecedented interpretability by explaining every generated token through three pathways: input context, human-understandable concepts, and training data provenance. Trained on 1.35 trillion tokens, it matches performance of models trained on 2-7x more data while enabling real-time concept control without retraining. This breakthrough addresses longstanding challenges in AI transparency and safety.

Built on a causal discrete diffusion model, Steerling-8B decomposes its outputs into 33,000 supervised concepts, 100,000 discovered concepts, and residual pathways. Over 84% of its token-level predictions flow through these interpretable components, allowing researchers to edit concepts at inference time. The model achieves 96.2% AUC in detecting known concepts, with ablation tests showing minimal performance loss when removing residual components.

Despite using fewer training FLOPs than LLaMA2-7B and Deepseek-7B, Steerling-8B outperforms both on standard benchmarks. Its architecture enables concept-level alignment through explicit steering mechanisms, replacing thousands of safety training examples with precise interventions. The model's weights and tools are publicly available via Hugging Face and PyPI.

This development marks a shift toward inherently interpretable AI systems, with potential applications in scientific research, regulatory compliance, and ethical AI deployment. Future work will explore discovered concepts, alignment without fine-tuning, and training data valuation.