Ornith-1.5 closes the self-improvement loop by generating its own tasks, scaffolds, and solution rollouts

For engineers building on open models, Ornith-1.5 suggests a path where the training data is no longer fixed but continuously generated by the model itself. This could reduce the need for human-curated task sets and hand-designed agent harnesses, though it also means the training distribution is shaped by the model's own capabilities and biases. The reported performance matches or exceeds larger closed models on agentic coding benchmarks, so the approach is worth watching.