Ornith-1.0 - DeepReinforce's Open-Source Agentic Programming Model Series

What is Ornith-1.0

Ornith-1.0 is an open-source large model series launched by the DeepReinforce team, specifically designed for Agentic programming tasks. The models are based on Gemma 4 and Qwen 3.5 pre-trained foundations, adopting a self-improving training framework, and achieving SOTA-level performance among open-source models of the same tier on code generation and software engineering benchmarks.

Main Features of Ornith-1.0

• Full-Spec Coverage: Offers four specifications: 9B Dense, 31B Dense, 35B MoE, and 397B MoE, meeting full-scenario deployment needs from edge devices to flagship servers.
• Agentic Programming: Optimized specifically for agentic programming tasks, supporting automated resolution of complex software engineering problems such as code repair, feature implementation, and terminal command execution.
• Self-Improving Scaffold Generation: The model can autonomously generate task-specific scaffolds (harness), guiding and optimizing the search path for solutions.
• Local Deployment Compatibility: Already available on HuggingFace, providing an OpenAI API-compatible interface and multiple local deployment options.

Follow WeChat and reply "Open Source" to join theAI Open Source Project Communication Group

Technical Principles of Ornith-1.0

• Foundation Models: Post-training is performed based on the pre-trained weights of Gemma 4 and Qwen 3.5, inheriting their language understanding and code generation capabilities.
• Self-Improving Training Framework: Unlike traditional RL that relies on manually designed scaffolds, Ornith-1.0 uses reinforcement learning to let the model jointly generate solution rollouts and task-specific harnesses, achieving collaborative optimization of the scaffold and the final solution.
• Joint Optimization Mechanism: During training, task scaffolds and generated results are optimized simultaneously, enabling the model to autonomously discover better search trajectories and continuously improve output quality.

How to Use Ornith-1.0

• Obtain the Model: Visit the HuggingFace platform and search for the Ornith-1.0 series to download the model weights of the corresponding specification.
• Local Deployment: Choose the 9B (edge devices), 31B/35B, or 397B version according to hardware configuration.
• API Calls: Use the OpenAI API-compatible interface to integrate into existing development workflows or Agent frameworks.
• Task Configuration: For Agentic coding scenarios, connect the model to IDE plugins or automated software engineering pipelines.

Core Advantages of Ornith-1.0

• Same-Tier SOTA: The 397B flagship version scored 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified, surpassing Claude Opus 4.7.
• Outperforming Higher Tiers: The 35B MoE version scored 64.4 on Terminal-Bench 2.1, surpassing the 397B-parameter Qwen 3.5-397B which scored 53.5, achieving more with less.
• Edge-Friendly: The 9B model can be deployed in low-resource environments, with performance matching Gemma 4-31B and Qwen 3.6 35B.

Ornith-1.0 Project Address

• Project Official Website：https://deep-reinforce.com/ornith_1_0.html
• HuggingFace Model Repository：https://huggingface.co/collections/deepreinforce-ai/ornith-10

Comparison of Ornith-1.0 with Similar Competitors

Application scenarios of Ornith-1.0