Ornith-1.5: From Self-Scaffolding to Self-Improvement

Ornith-1.5 extends the Ornith-1.0 self-scaffolding system into a full self-improvement loop that autonomously creates tasks, builds task-specific scaffolds, and generates solution rollouts for reinforcement learning. It is released in 397 B MoE, 35 B MoE, and 9 B dense versions, matches Claude Opus 4.8 on Terminal-Bench 2.1 (86.1) and DeepSWE (56), and exceeds comparable open-source models on reasoning, coding, and agentic benchmarks; the 9 B model runs on mobile devices and outperforms larger models such as Gemma 4-31 B and Qwen 3.6-35 B.