Open-Source Coding Model Ornith-1.0 Writes Its Own Training Scaffold in Reinforcement Learning
DeepReinforce released Ornith-1.0 on June 25, 2026 — a family of four MIT-licensed open-source coding models where the model learns to generate its own training scaffold during reinforcement learning, rather than relying on a fixed human-designed harness.
Model Lineup
| Model | Architecture | Target |
|---|---|---|
| Ornith-1.0-9B | Dense | Edge / single 80GB GPU (~19GB BF16) |
| Ornith-1.0-31B | Dense | General-purpose mid-tier |
| Ornith-1.0-35B | MoE | Strong performance at lower inference cost |
| Ornith-1.0-397B | MoE flagship | Maximum capability; FP8/GGUF builds available |
All models are post-trained on Gemma 4 and Qwen 3.5 foundations and are reasoning models by default (chain-of-thought before final answer).
Self-Scaffolding RL
Each RL step runs in two stages: (1) the model proposes a refined scaffold for the task, (2) generates a solution conditioned on that scaffold. Reward propagates to both stages, so harness design and code quality improve together.
Reward hacking is mitigated via a three-layer defense: immutable trust boundary (environment/tools), deterministic monitor (zero reward for boundary violations), and frozen LLM judge veto.
Benchmark Claims
DeepReinforce reports for the 397B flagship:
- 82.4 on SWE-Bench Verified (vs Claude Opus 4.7 at 80.8)
- 77.5 on Terminal-Bench 2.1 (vs Opus 4.7 at 70.3)
- 62.2 on SWE-Bench Pro
The team claims state-of-the-art among open-source models of comparable parameter count, not overall leaderboard leadership (Claude Opus 4.8 scores higher on several benchmarks).
Smaller models show efficiency gains: 35B MoE scores 64.2 on Terminal-Bench 2.1 vs Qwen 3.5-397B at 53.5; 9B dense reaches 43.1 on Terminal-Bench 2.1.
Context and Caveats
DeepReinforce’s prior work includes CUDA-L1 and IterX. Independent research has documented SWE-Bench Verified limitations including semantic incorrectness in ~20% of “resolved” patches and solution leakage in 32%+ of instances. Community sentiment on launch day was mixed (55.4% negative in tracked social engagement).
Weights and deployment recipes are on Hugging Face under MIT license.