DeepReinforce released Ornith-1.0 on June 25 under MIT license, a family of open-source coding models specialized for agentic AI tasks. Available in four sizes (9B, 31B, 35B MoE, 397B MoE), the models excel at autonomous multi-step coding workflows without human intervention. The 397B flagship scores 82.4 on SWE-bench Verified, beating Claude Opus 4.7 (80.8) and DeepSeek-V4-Pro (80.6). On Terminal Bench 2.1, it achieves 77.5 vs Claude's 70.3. The 9B variant scores 69.4, outperforming Google's Gemma 4-31B (52.0) and nearly matching Qwen 3.5-35B (70.0). Ornith uses a novel training method where the model learns its own task decomposition strategy via reinforcement learning, co-evolving with the scaffold. However, the model card warns it may underperform on non-coding tasks, making it unsuitable for general-purpose use. DeepReinforce also implements defenses against reward hacking, including immutable environments and a deterministic monitor.

Key facts

• Ornith-1.0 is open-source under MIT license, available in 9B, 31B, 35B MoE, and 397B MoE sizes.
• 397B flagship scores 82.4 on SWE-bench Verified, surpassing Claude Opus 4.7 and DeepSeek-V4-Pro.
• 9B variant scores 69.4 on SWE-bench, outperforming Gemma 4-31B (52.0) and near Qwen 3.5-35B (70.0).
• Model learns its own task strategy via reinforcement learning, not relying on human-designed scaffolds.
• Not suitable for general tasks; optimized solely for agentic coding in developer pipelines.

KeyAudit data perspective