Ornith-1.0 provides a suite of open-source, self-improving models specifically designed for agentic coding tasks. Targeting developers and researchers, it offers state-of-the-art performance on various coding benchmarks, enabling more sophisticated AI-assisted code generation and manipulation. The project's core innovation lies in its reinforcement learning framework, which enhances agentic capabilities.

How It Works

Ornith-1.0 employs a novel self-improving training framework that uses Reinforcement Learning (RL) to optimize not only the generation of code solutions but also the underlying "scaffolds" or search trajectories that lead to those solutions. By jointly optimizing both aspects, the model discovers more effective problem-solving strategies, resulting in higher-quality code outputs. The models are built upon foundational architectures like Gemma 4 and Qwen 3.5.

Quick Start & Requirements

Serving Ornith-1.0 requires recent runtimes: Transformers ≥ 5.8.1, vLLM ≥ 0.19.1, or SGLang ≥ 0.5.9. The 9B dense model fits on a single 80GB GPU, while larger 35B and 397B Mixture-of-Experts (MoE) models require multi-GPU nodes with tensor parallelism. All models support a 256K (262,144-token) context window. Variants like FP8 and GGUF are available for reduced VRAM usage and local inference via llama.cpp/Ollama. Serving examples are provided for vLLM and SGLang, alongside direct Hugging Face Transformers integration.

Highlighted Details

• Achieves state-of-the-art performance on coding benchmarks including Terminal-Bench 2.1, SWE-Bench, NL2Repo, and OpenClaw.
• Available in multiple sizes: 9B-Dense, 35B-MoE, and 397B-MoE, post-trained on Gemma 4 and Qwen 3.5.
• Features a self-improving training framework using RL for enhanced agentic reasoning and solution generation.
• Exposes an OpenAI-compatible API endpoint, facilitating integration with various agent frameworks and SDKs.
• Supports an extensive 256K token context window.

Maintenance & Community

The project is developed by the "DeepReinforce Team." Further details and updates can be found on their blog. [INDEX]

Licensing & Compatibility

Ornith-1.0 is MIT licensed, ensuring global accessibility and freedom from regional restrictions. Its OpenAI-compatible API makes it broadly compatible with existing agent frameworks, LLM clients, and coding CLIs.

Limitations & Caveats

The README does not explicitly detail limitations. However, deploying the larger MoE models necessitates significant multi-GPU infrastructure. As a research-oriented project, ongoing maintenance and community support structures may differ from commercial offerings.