Ornith-1.0-35B-FP8

Ornith-1.0-35B-FP8 is a compact variant within the Ornith family, designed for efficient single-GPU deployment while preserving the agentic capabilities of larger siblings. It belongs to a family of open-source models for agentic coding, available in 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE configurations, post-trained on top of Gemma 4 and Qwen 3.5 foundations. The model employs a self-improving training framework that uses reinforcement learning to jointly optimize both the scaffold driving rollouts and the resulting solutions, helping the model discover stronger search trajectories and produce higher-quality code-related outputs.

This 35B variant delivers benchmark results that place it ahead of comparable open-source peers on several agentic coding evaluations. It scores 64.2 on Terminal-Bench 2.1 Terminus-2 and 62.8 on the Claude Code variant, while reaching 75.6 on SWE-bench Verified, 50.4 on SWE-bench Pro, and 69.3 on SWE-bench Multilingual. It also posts 34.6 on NL2Repo. Its lightweight footprint makes it a practical fit for teams wanting single-GPU serving without giving up strong performance on real software-engineering tasks, and the MIT-licensed release keeps it globally accessible for open experimentation and deployment.