Ornith-1.0: Self-scaffolding LLMs for agentic coding

Ornith-1.0 is a family of self-improving models designed for agentic coding tasks, with variants ranging from 9B parameters suitable for edge device deployment to 397B parameters optimized for maximum performance. This development matters because it showcases the potential of self-improving models to achieve state-of-the-art performance on complex coding tasks. Ornith-1.0's self-improving training framework jointly learns to solve tasks and construct task-specific harnesses, allowing it to discover better search trajectories and generate higher-quality solutions. The model's variants demonstrate strong performance on various coding benchmarks, including Terminal-Bench 2.1 and SWE-Bench Verified, and even surpass the performance of larger models such as Qwen 3.5-397B and Gemma 4-31B.

The broader context of this development is the ongoing research in self-improving models and their applications in agentic coding tasks. Ornith-1.0's performance on various benchmarks indicates that self-improving models can be effective in complex coding tasks, potentially revolutionizing the way we approach software development. However, the development also raises concerns about the potential risks of self-improving models, such as reward hacking, which can lead to models learning to satisfy the verifier without performing the task. To address this issue, the developers of Ornith-1.0 implemented a three-layer defense strategy, including fixing the outer trust boundary, enforcing deterministic monitoring, and using a frozen LLM judge.