Ornith Releases Version 1.5

Ornith has released Ornith version 1.5, introducing a new family of open-weight models built around an expanded self-improvement loop. The release spans three distinct model sizes: a 397B mixture-of-experts flagship, a 35B mixture-of-experts model that activates 3 billion parameters per token, and a 9B dense model complete with a quantized Mobile build for iPhone and Android devices.

The Self-Improvement Loop

Building upon the self-scaffolding framework from the previous generation, Ornith-1.5 automates the training cycle by having the model propose new tasks, generate task-specific evaluation harnesses, and produce solution rollouts for reinforcement learning. The system rewards tasks based on validity, frontier difficulty — targeting a 0.2 empirical success rate to maintain an appropriate challenge — and novelty, using validity as a hard gate to filter out malformed environments.

Model Scales and Benchmarks

According to published laboratory results averaged over five independent runs, the flagship 397B model achieves strong scores on developer benchmarks, such as 86.1 on Terminal-Bench 2.1 and 56 on DeepSWE. The mid-sized 35B model and the compact 9B model also deliver competitive performance for their parameter counts, making local deployment and specialized agentic tasks more accessible to developers working outside of closed cloud environments.

Verified details

Ornith has released Ornith version 1.5.