deepreinforce-ai/ornith-1

Ornith-1.0 is a collection of open-source AI models built specifically for agentic coding, meaning they are designed to autonomously write, fix, and manage code in real software projects, not just answer one-off questions. The project offers several model sizes (9B, 31B, 35B, and 397B parameters), and the team reports that these models outperform other open-source models of similar size on well-known coding benchmarks like SWE-Bench and Terminal-Bench. What makes Ornith interesting is its "self-improving" training approach. Instead of just training the model to produce code solutions, the training process also teaches the model to generate its own scaffolding, the step-by-step structure and search strategy that guides how it arrives at a solution. By learning to improve both the problem-solving approach and the code itself simultaneously, the model gets better at finding effective paths to correct answers. The benchmark results in the README compare Ornith against other open models from the Qwen and Gemma families. For example, the 9B model scores 69.4 on SWE-bench Verified, compared to 53.2 for Qwen3.5-9B. The 35B model reaches 75.6 on the same benchmark, beating Qwen3.5-35B's score of 70. These benchmarks test real-world tasks like fixing bugs in open-source repositories and navigating terminal-based coding challenges. This project would appeal to developers, startups, or teams that want a strong open-source coding model they can run or fine-tune themselves, rather than relying on closed proprietary APIs. The MIT license means there are no regional restrictions or commercial usage barriers, which matters for teams operating globally or building products on top of the model. The README is heavily focused on benchmark tables and doesn't go into detail about hardware requirements, deployment instructions, or practical setup guidance, so prospective users would need to look elsewhere for that information.