DeepReinforce Open-Sources the Ornith-1.0 Programming Model: 35B MoE Outperforms Larger Models

ME AI News: According to monitoring by Beating, the AI team DeepReinforce, founded by renowned NLP scholar Li Jiwei, has open-sourced four Ornith-1.0 models specialized for programming agents, all released under the MIT license. The Ornith-1.0 models are trained on Gemma 4 and Qwen 3.5, featuring both Dense and MoE (Mixture of Experts) architectures. In the Terminal-Bench 2.1 (Claude Code) programming agent benchmark, the flagship Ornith-1.0-397B MoE achieved a score of 78.2, surpassing the closed-source model Claude Opus 4.7 at 69.7. The core innovation of Ornith-1.0 lies in jointly training the execution scaffold with the code policy network via reinforcement learning. Unlike traditional fixed scaffolds, the model dynamically evolves and updates its scaffold—comprising memory, error handling, and orchestration logic—at each training step before generating code solutions based on the new scaffold, with reward signals backpropagated to both stages. To prevent the model from cheating by reading hidden test files or hardcoding answers during self-generated scaffold creation, the team implemented a deterministic monitor with interception and a LLM judge veto mechanism, establishing a three-layer defense system. Evaluation data shows that medium and lightweight models can outperform larger ones. In the same Claude Code testing environment, the 35B MoE model Ornith-1.0-35B MoE achieved a score of 62.8, surpassing the larger 397B parameter model Qwen 3.5-397B (48.6) and Qwen 3.6-35B (49.2). The end-device-optimized Ornith-1.0-9B Dense model also achieved 69.4 on the SWE-Bench Verified test, outperforming the 31B-scale Gemma 4-31B (52.0). (Source: BlockBeats)