Ornith-1.5模型发布：自生成训练任务，自测编码得分超Opus 4.8
Ornith-1.5让AI在训练中自己生成题目并求解，自测编码能力超越闭源旗舰，但成绩未获官方验证榜确认。
Ornith-1.5用强化学习让模型自己出题、自己搭解题环境、自己跑解答轨迹，Terminal-Bench自测86.1分，超过Claude Opus 4.8的85.0。但该成绩为项目自测，未出现在官方验证榜单上。同时发布的35B和9B小模型，在低算力下表现亮眼。
正文摘录
 AI Frontier 报道  To train a model that actually works, the most expensive thing isn't the GPUs — it's the training problems. Especially when you get to coding and agentic models. Having questions and answers isn't enough — you also need a runnable environment for each problem, a test that can judge right from wrong, and protection against the model rummaging through old submissions to copy answers. The difficulty can't be random either. Too easy and the model finishes with its eyes closed, learning nothing; too hard and it can't solve a single one, accumulating no correct experience — a wasted round of compute. And even once you have all…