New ReleaseJune 29, 2026 · 5 min read
🐦 Ornith 1.0: Self-Scaffolding Coding Agents
A new family of open-source coding models learns to write its own agent harness during training — and the 397B variant beats Claude Opus 4.7 on Terminal-Bench.
📋 Article Summary
- 🔑 Self-Scaffolding Explained — Model jointly learns harness + solution, not just answers
- 🏗️ Two-Stage RL — Scaffold generation → solution rollout → GRPO reward to both
- 🛡️ Anti-Reward-Hacking — Immutable env + deterministic monitor + LM judge veto
- 📊 Benchmarks — 397B: 82.4% SWE-bench, 77.5% Terminal-Bench. 9B: edge-deployable
- 💡 Why It Matters — First framework where orchestration is learned, not hand-engineered
The Innovation
DeepReinforce's Ornith-1.0 is a family of four models (9B Dense, 31B Dense, 35B MoE, 397B MoE) fine-tuned from Qwen 3.5 and Gemma 4. The breakthrough isn't the base architecture — it's the training method: the model learns to write its own agent harness during reinforcement learning.
How It Works
- Scaffold Generation — Model reads task + previous scaffold, proposes refined harness.
- Solution Rollout — Uses that scaffold to generate solution. Reward flows back to both stages.
- GRPO Optimization — Updates weights for both scaffold and solution quality.
Anti-Reward-Hacking
- • Immutable environment — Sandbox tools can't be modified by model
- • Deterministic monitor — Watches for unauthorized tool use, immediate penalty
- • LM judge veto — Can override automated checks if solution seems suspicious
Benchmark Results
| Model | SWE-Bench | Terminal-Bench | Params |
|---|---|---|---|
| Ornith 397B | 82.4 | 77.5 | 397B MoE |
| Ornith 35B | 75.6 | 64.4 | 35B MoE |
| Ornith 9B | 69.4 | 43.1 | 9B Dense |
| Claude Opus 4.7 | 80.8 | 70.3 | ~500B+ |
| DeepSeek V4 Pro | 80.6 | 67.9 | 1.6T/49B |
Key takeaway: The 397B variant beats Claude Opus 4.7 on Terminal-Bench (77.5 vs 70.3) and matches it on SWE-bench. The 9B punches well above its weight at 69.4% SWE-bench. All MIT-licensed with open weights.