Ornith-1.0: A family of open-source LLMs specialized for agentic coding (twitter.com)

Ornith-1.0 has been unveiled as a new family of open-source large language models (LLMs) specifically designed for agentic coding, encompassing various sizes from 9B to a whopping 397B parameters. This release is notable as it achieves state-of-the-art performance on several coding benchmarks, outperforming other models in its class, including Terminal-Bench and SWE-Bench. These benchmarks illustrate the model's robust capabilities in tackling complex coding tasks and generating high-quality solutions. The significance of Ornith-1.0 lies in its innovative training method, which incorporates reinforcement learning not only for generating solution rollouts but also for creating task-specific scaffolds that enhance those rollouts. This dual optimization approach leads to improved performance in coding tasks, making it a valuable tool for developers and researchers alike. Released under the MIT license, Ornith-1.0’s models are freely available for both commercial and research purposes, ensuring broad accessibility within the AI/ML community. This positions Ornith-1.0 as a potentially transformative resource for advancing the capabilities of coding AI systems.