Ornith-1.0, a coding AI model with performance equivalent to Claude Opus 4.7, has been released, and a smaller, locally-running model is also available.



DeepReinforce , an AI research organization, has released its AI model family ' Ornith-1.0 ' as an open model. The top-of-the-line model, Ornith-1.0-397B, has outperformed Claude Opus 4.7 in multiple benchmark tests.

Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding | DeepReinforce Blog | Jun. 2026
https://deep-reinforce.com/ornith_1_0.html




Ornith-1.0 is an AI model specifically designed for agent coding, developed based on Gemma 4 and Qwen 3.5. While scaffolds and harnesses that determine the behavior of agent AIs are often built by humans, Ornith-1.0 also learns the scaffold, successfully optimizing it while avoiding reward hacking and thereby improving the overall performance of the agent AI.



Four versions of Ornith-1.0 have been announced: 'Ornith-1.0-9B,' 'Ornith-1.0-31B,' 'Ornith-1.0-35B,' and 'Ornith-1.0-397B.' Ornith-1.0-9B and Ornith-1.0-31B are Dense models that read and process all parameters, while Ornith-1.0-35B and Ornith-1.0-397B are MoE models that activate only some parameters depending on the input. The largest model, Ornith-1.0-397B, is comparable to Claude Opus 4.7, while the smallest model, Ornith-1.0-9B, is positioned as a model that operates on edge devices.

The graph below compares the benchmark scores of the Ornith-1.0-397B with models of similar size. The Ornith-1.0-397B outperformed the Claude Opus 4.7 in several tests.



Ornith-1.0-35B also outperforms Qwen3.6 35B and Gemma 4 31B.



Furthermore, even the smallest Ornith-1.0-9B outperformed the relatively large Gemma 4 31B.



The three versions, 'Ornith-1.0-9B,' 'Ornith-1.0-35B,' and 'Ornith-1.0-397B,' are available at the following link. The license is the MIT License . Ornith-1.0-31B will also be released soon.

Ornith-1.0 - a deepreinforce-ai Collection
https://huggingface.co/collections/deepreinforce-ai/ornith-10



Related Posts:

in AI, Posted by log1o_hf