DeepReinforce releases open-source coding model Ornith-1.0

DeepReinforce has released Ornith-1.0, a self-learning model family designed for Agentic Coding, as open source. The lineup ranges from a 9B Dense model for edge devices to a 397B MoE model for state-of-the-art tasks, and was trained on top of Gemma 4 and Qwen 3.5.

The key differentiator of Ornith-1.0 is that the model generates not just the solution but also the Scaffold that guides it. Instead of relying on human-designed scaffolds, during the RL (Reinforcement Learning) stage the model proposes an optimal scaffold and generates a solution based on it, learning orchestration ability alongside the reward obtained through this process.

To prevent Reward Hacking that could occur when the model writes its own scaffolds, three layers of defense were introduced:

• A fixed external trust boundary that isolates the environment and tests
• A deterministic monitor that detects modification of verification scripts
• A fixed LLM judge that blocks cheating during tool use

In terms of performance, the 397B model scored 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, showing performance on par with Claude Opus 4.7. Notably, the 9B model delivers coding ability comparable to larger models like Gemma 4-31B even on lower-spec hardware.