DeepReinforce AI released Ornith-1.0, a set of self-improving open-source models designed for agentic coding tasks. The release includes four variants: 9B-Dense, 31B-Dense, 35B-Mixture-of-Experts, and 397B-Mixture-of-Experts. All models are post-trained on Gemma 4 and Qwen 3.5 foundations, targeting developers who need automated code generation and problem-solving capabilities.

What sets Ornith-1.0 apart is its reinforcement learning framework that simultaneously optimizes both solution rollouts and their underlying scaffolds. This dual optimization approach helps the model discover better search trajectories and produce higher-quality code solutions. The 397B variant scores 77.5 on Terminal-Bench 2.1, outperforming several proprietary models in its size class. Benchmark results show consistent improvements across SWE-bench Verified, NL2Repo, and Claw-eval tests.

The project uses an MIT license with no regional restrictions, making it globally accessible for commercial and research use. Each model variant ships in multiple formats including bf16, FP8, and GGUF quantization. The dense 9B fits on a single 80GB GPU while larger MoE versions require multi-GPU setups with tensor parallelism.

Ornith-1.0 exposes an OpenAI-compatible API interface and supports 256K context windows. Developers can deploy via vLLM, SGLang, or Hugging Face Transformers with recommended sampling parameters of temperature 0.6 and top_p 0.95. The models output reasoning traces in dedicated fields, enabling integration with existing agentic coding workflows.