A minimal, stripped-down model built for testing infrastructure rather than production use. It produces text embeddings from short inputs, with a 512-token context window that keeps scope tightly bounded. Expect limited capability — this is a randomly initialized DeBERTa variant, meaning weights carry no learned knowledge.