Core Mechanism

• The Ornith 1.0 models—built as fine-tuned variants of Qwen 3.5 and Gemma 4—utilize a two-stage reinforcement learning loop that trains the model to first refine a task-specific scaffold and then generate a solution rollout conditioned on that scaffold, updating the policy based on rewards from both stages.2:18
• This self-scaffolding approach aims to replace manual context engineering, allowing the model to adapt its own tooling dynamically; for instance, the model successfully swapped an OpenWeather API requirement for an open-source Open-Meteo alternative when no keys were provided.3:44

Safety and Performance

• To prevent reward hacking—where models exploit sandbox vulnerabilities to maximize scores—the system employs three layers: immutable tools, a deterministic monitor for forbidden actions, and an LLM-judge veto.7:06
• While the speaker reports that the largest 397B MoE model is competitive with top-tier systems like Claude Opus, the transcript provides no raw benchmark data to support these findings.2:43
• The project is fully open to the public; the speaker highlights the 9B model as a viable option for local evaluation on hardware lacking significant GPU resources.1:55