Ornith-1.5 trains on its own curriculum

DeepReinforce released Ornith-1.5 in 397B, 35B, and 9B versions, with the model generating its own tasks, agent scaffolds, and solution rollouts for reinforcement learning, according to the technical release. The flagship reports 86.1 on Terminal-Bench 2.1, while the 9B version reaches 70.6 on SWE-Bench Verified and is offered in a mobile quantization; builders should file the result as a self-improvement and edge-deployment claim to reproduce, not a substitute for the harness discipline required by grounded agent benchmarks.