Ornith 1.0 FAQ: Everything You Need to Know

Ornith 1.0 is a family of open-source large language models built by DeepReinforce AI specifically for agentic coding. Ornith 1.0 comes in four sizes — 9B Dense, 31B Dense, 35B MoE, and 397B MoE — all released under MIT license. The key innovation of Ornith 1.0 is self-scaffolding: the model jointly learns to solve coding tasks and construct the orchestration framework that guides those solutions.

Ornith 1.0 was created by DeepReinforce AI and released on June 25, 2026. The name comes from the ancient Greek word for bird. All Ornith 1.0 models are available on Hugging Face under the deepreinforce-ai organization with MIT licensing and no regional restrictions.

Ornith 1.0 models are built on two base architectures. The 9B Dense, 35B MoE, and 397B MoE variants are post-trained on Qwen 3.5. The 31B Dense variant is post-trained on Gemma 4. All variants undergo the same self-improving reinforcement learning process that jointly optimizes scaffolds and solutions.

For most users, Ornith 1.0-35B MoE is the sweet spot — it is actually faster than the 9B model due to MoE architecture (only ~3B parameters active per token) while being significantly more accurate. If you only have 6-8GB VRAM, Ornith 1.0-9B Q4 is a realistic entry point. The 397B model is for production agent pipelines where maximum accuracy matters.

Ornith 1.0-9B needs about 6 GB in Q4 quantization or 19 GB in bf16. Ornith 1.0-35B MoE needs about 25 GB in Q5_K_M. Ornith 1.0-397B requires approximately 200 GB in FP8 or 400 GB in bf16, typically served across 8x 80GB GPUs. The 35B MoE is the best option for consumer GPUs with 24GB+ VRAM.

Traditional coding agents use a human-designed scaffold — a fixed harness that wraps the model with memory, tools, and error handling. Ornith 1.0 instead treats the scaffold as a learnable object. During RL training, the scaffold co-evolves with the model's policy, meaning Ornith 1.0 learns to generate its own task plans, tool calls, error recovery, and re-planning strategies.

Ornith 1.0-397B scores 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, surpassing Claude Opus 4.7 (70.3 and 80.8 respectively) but trailing Claude Opus 4.8 (85 and 87.6). The key advantage of Ornith 1.0 is that it is fully open-source under MIT license, so you can self-host and run it with zero API costs.

Yes. Ornith 1.0 exposes an OpenAI-compatible endpoint when served via vLLM, SGLang, or llama.cpp. It works directly with Claude Code, OpenHands, OpenClaw, Hermes Agent, OpenCode, and any other tool that supports the OpenAI API format. No code changes needed — just point your tool's base URL at your Ornith 1.0 server.

Yes. All four Ornith 1.0 model variants are released under the MIT license with no regional restrictions. The weights are freely downloadable from Hugging Face, and GGUF quantized versions are also provided for local deployment on consumer hardware. You can use Ornith 1.0 for any purpose including commercial applications.

Ornith 1.0 uses three layers of anti-reward-hacking safeguards during RL training: a fixed trust boundary that limits what the model can access, a deterministic monitor that flags suspicious patterns, and a frozen LLM judge that evaluates solution quality independently of the reward signal. This ensures benchmark scores reflect genuine capability.