Ornith-1.0: self-improving open-source models for agentic coding (github.com)

The launch of Ornith-1.0 introduces advanced self-improving open-source models for agentic coding, which are available in configurations including 9B-Dense, 31B-Dense, 35B-MoE, and a robust 397B-MoE variant. These models, which have been fine-tuned on Gemma 4 and Qwen 3.5, outperform existing open-source models in coding benchmarks like Terminal-Bench 2.1 and SWE-bench, marking a significant advancement in AI-driven coding capabilities. This performance enhancement is achieved through an innovative self-improving training framework that leverages reinforcement learning (RL) to optimize not just the solutions but also the scaffolding used to derive them, enabling the discovery of more efficient search paths and higher-quality outcomes. Ornith-1.0's implications for the AI/ML community are substantial, as it democratizes access to sophisticated coding agents with its MIT license and global availability. Each model is designed to be compatible with existing OpenAI tools and frameworks, supporting a wide context window of 256K tokens. This flexibility, coupled with a clear focus on tool-calling and agentic functionality, positions Ornith-1.0 as a powerful resource for developers and researchers, allowing for enhanced interactions with code generation and manipulation tasks within various AI applications.