DeepReinforce Releases Open Coding AI Ornith-1.0, Top Model Beats Claude Opus 4.7
Ornith-1.0 treats agent scaffolds as learning targets rather than human-designed components, and its smallest model outperforms a much larger competitor, suggesting a shift in how efficient coding agents can be built.
Reporting from 1 sources: GIGAZINE.
AI research organization DeepReinforce has released the Ornith-1.0 family of open models specialized for agent coding. The largest variant, Ornith-1.0-397B, outperforms Claude Opus 4.7 on multiple benchmarks. Four variants range from 9B to 397B parameters, with the smallest designed to run on edge devices. The models are distributed under the MIT License.
DeepReinforce introduced Ornith-1.0, a family of open AI models built for agent coding. The largest model, Ornith-1.0-397B, beat Claude Opus 4.7 across several benchmarks. Unlike typical agent systems where humans design the scaffold that controls behavior, Ornith-1.0 learns the scaffold as part of training while avoiding reward hacking. Four variants were announced: Ornith-1.0-9B, Ornith-1.0-31B, Ornith-1.0-35B, and Ornith-1.0-397B. The 9B and 31B are Dense models; the 35B and 397B use Mixture-of-Experts architecture. The 9B model outperformed the much larger Gemma 4 31B. Three variants are available now under the MIT License, with the 31B coming soon.
- Ornith-1.0-9B: Dense model, runs on edge devices, outperformed Gemma 4 31B
- Ornith-1.0-31B: Dense model, scheduled for release soon
- Ornith-1.0-35B: MoE model, outperformed Qwen3.6 35B and Gemma 4 31B
- Ornith-1.0-397B: MoE model, comparable to Claude Opus 4.7, outperformed it on multiple benchmarks
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.