DeepReinforce open-sourced Ornith-1.0, a family of coding models that generates its own reinforcement learning scaffolds and matches Claude Opus 4.7 on benchmarks.

The release spans four model sizes, from a 9B Dense version for edge deployment to a 397B mixture-of-experts model for frontier-scale work. The lineup includes 31B Dense and 35B MoE variants, all built on pretrained Gemma 4 and Qwen 3.5 foundations.

Self-scaffolding approach

Ornith-1.0's key innovation lies in how it handles reinforcement learning scaffolds. Rather than relying on human-designed harnesses, the model learns to produce both solution rollouts and the task-specific scaffolds that guide them.

Each RL step runs in two stages: the model first proposes a refined scaffold conditioned on the task, then generates a solution using that scaffold. Reward signals flow back to both stages, training the model to author both orchestration and answers.

This approach lets scaffolds mutate and evolve toward higher-reward trajectories without hand-engineered harness design.

Security measures

To prevent reward hacking, DeepReinforce implemented three defensive layers:

  • A fixed outer trust boundary that isolates the environment and tests from model access
  • A deterministic monitor flagging attempts to read restricted paths or alter verification scripts
  • A frozen LLM judge that vetoes the verifier when gaming occurs within allowed tool surfaces

Performance benchmarks

The 397B flagship model achieved 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, matching Claude Opus 4.7 and competing with MiniMax M3 and DeepSeek-V4-Pro.

The 9B version scored 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified, matching much larger models like Gemma 4-31B.

DeepReinforce previously shipped CUDA-L1 research and the IterX optimization loop for code agents. The company released Ornith-1.0 weights and technical documentation on Hugging Face for direct deployment and research use.