Ornith 1.0 397B
deepreinforce-ai/Ornith-1.0-397B
published Jun 2026 · updated Jun 2026
Ornith 1.0 397B is a text-generation model that achieves state-of-the-art performance in agentic coding tasks using a self-improving reinforcement learning framework.
specs
| Task | Text Generation |
| Architecture | Mixture of Experts (MoE) |
| Parameters | 397B |
| License | MIT |
about this model
Key Strengths
The model employs a self-improving training framework that uses reinforcement learning to jointly optimize both the scaffold (the search or tool-use strategy) and the solution rollouts. This approach enables the model to discover better search trajectories and generate higher-quality solutions. A three-layer defense mechanism protects against reward hacking: a fixed outer trust boundary with immutable environment/tool/test isolation, a deterministic monitor that flags attempts to read withheld paths or modify verification scripts, and a frozen LLM judge for intent-level gaming detection.
Benchmark Performance
Ornith-1.0-397B achieves state-of-the-art results among open-source models of comparable size on agentic coding benchmarks:
| Benchmark | Ornith-1.0-397B | Qwen3.5-397B | Qwen3.7-Max | DeepSeek-V4-Pro-1.6T | Claude Opus 4.7 |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 (Terminus-2) | 77.5 | 53.5 | 73.5 | 64 | 70.3 |
| Terminal-Bench 2.1 (Claude Code) | 78.2 | 48.6 | 69.8 | 66.5 | 69.7 |
| SWE-bench Verified | 82.4 | 76.4 | 80.4 | 80.6 | 80.8 |
| SWE-bench Pro | 62.2 | 51.6 | 60.6 | 55.4 | 64.3 |
| SWE-bench Multilingual | 78.9 | 69.3 | 78.3 | 76.2 | - |
| NL2Repo | 48.2 | 36.8 | 47.2 | - | - |
| Claw-eval Avg | 77.1 | 70.7 | 65.2 | 75.8 | 78.2 |
| SWE Atlas - QnA | 41.2 | 20.4 | - | 27.2 | 40.3 |
| SWE Atlas - RF | 42.6 | 18.4 | - | 27.2 | 40.3 |
Additional Capabilities
The model supports vision and video inputs via dedicated tokens in its chat template, and includes a built-in tool-calling template with XML-based function calling format. It is MIT licensed.
best for
- ·Automated bug fixing and patch generation in large codebases
- ·End-to-end repository-level code generation from natural language descriptions
- ·Building autonomous coding agents for multi-step software engineering tasks
FAQ
It excels at agentic coding tasks such as SWE-Bench, Terminal-Bench, and repository-level code generation, making it ideal for automated software engineering.
It achieves state-of-the-art results among open-source models of comparable size, outperforming Qwen 3.5-397B and matching or exceeding larger proprietary models on several benchmarks.
The model is released under the MIT license, globally accessible and free from regional limitations.
Yes, it supports vision and video inputs via special tokens (e.g., <|vision_start|>) in its chat template.
Use the gigarouter OpenAI-compatible endpoint with your API key to send prompts and receive generated text.
We're benchmarking and onboarding Ornith 1.0 397B as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.