DeepReinforce Releases Ornith-1.0: An Open-Source Coding Model Family That Learns Its Own RL Scaffolds

An open-source coding model family called Ornith-1.0 has been released in four sizes, ranging from 9 billion to 397 billion parameters. Unlike typical coding agents that use a fixed human-designed framework, Ornith-1.0 learns to write and refine its own framework during training, optimizing both the framework and the solution together. The models are designed to handle coding tasks like multi-file refactors and bug fixes, and the largest version performs competitively with other open models on coding benchmarks. The release includes safeguards against reward hacking through fixed trust boundaries, action monitoring, and a frozen judge to prevent the model from gaming the reward system.