Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding (deep-reinforce.com)

Ornith-1.0 has been unveiled as a groundbreaking family of self-improving open-source models designed for agentic coding tasks, featuring a range of options from compact 9B Dense models for edge deployment to a powerful 397B MoE variant. The models leverage a self-improving training framework that not only learns to generate solutions but also develops the task-specific scaffolds to optimize those solutions. This innovative approach allows Ornith-1.0 to achieve state-of-the-art performance on various coding benchmarks, boasting impressive scores like 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, surpassing competitive models such as Claude Opus 4.7 and MiniMax M3. The significance of Ornith-1.0 lies in its ability to adaptively construct scaffolds, thereby reducing dependency on human-led design while promoting enhanced learning trajectories. To counter reward-hacking risks associated with self-generated scaffolding, Ornith-1.0 employs a multi-layered defense strategy, including immutable environment constraints and a deterministic monitoring system. This model's adaptation to edge deployment, particularly with the 9B variant outperforming much larger counterparts, signals a promising advancement in delivering high-performance AI capabilities across diverse computing environments.