Instead of relying only on fixed human-curated tasks and hand-built harnesses, Ornith-1.5 proposes new tasks, generates task-specific scaffolds, produces solution rollouts, and uses reinforcement learning to improve its policy. This creates new learning experiences continuously as the system trains.
The 9B model and a quantized mobile version target iPhone and Android deployment, while the larger models target demanding engineering and research workloads. Ornith-1.5 is useful for coding agents, self-improving training systems, edge experimentation, and research into automated task and scaffold generation.
