Ornith-1.0: Open-source coding models for agents

DeepReinforce released Ornith-1.0 on June 25. The lab published the models under an MIT license and posted weights and documentation on Hugging Face. Ornith-1.0 is intended to operate inside terminals and code repositories for automated developer workflows rather than as a conversational assistant.

The family includes four sizes: a 9 billion-parameter dense model, a 31 billion-parameter dense model, a 35 billion-parameter mixture-of-experts model and a 397 billion-parameter mixture-of-experts flagship. DeepReinforce described the release as “a self-improving family of open-source models specially for agentic coding tasks.” The project follows the lab’s earlier work on CUDA-L1 and the IterX loop for optimizing code agents.

Agentic coding means the model takes actions inside a development environment: reading files, running tests, diagnosing failures, editing code and iterating until a task is complete. Ornith is trained and evaluated to perform these multi-step workflows autonomously inside repositories and containerized terminals.

During reinforcement learning, training steps run in two phases. The model first proposes a strategy for a task and then executes that strategy. Rewards from task outcomes are applied to both strategy generation and execution, so the model learns to write better plans as well as better code.

DeepReinforce built several safeguards to limit reward hacking, where a model could try to appear to pass verification without solving the underlying problem. The protections include immutable test suites outside the model’s control, a deterministic monitor that detects attempts to access restricted files or change verification scripts, and a frozen judge model that can veto suspicious verifier results.

On public benchmarks, the 397B model scored 82.4 on SWE-bench Verified, a test that gives models real bugs from open-source GitHub repositories and scores successful fixes. The 397B scored 62.2 on SWE-bench Pro, a variant designed to reduce benchmark contamination. On Terminal Bench 2.1, which runs 89 tasks in containerized terminals, the 397B recorded a 77.5 percent completion rate. The 9B model scored 69.4 on SWE-bench Verified; that result is higher than Gemma 4-31B’s 52 and close to reported results for other models near 70 on the same test.

The model card for Ornith-1.0 warns the family may underperform on tasks outside agentic coding, including document summarization, long-form academic writing and general conversation. DeepReinforce positions Ornith for teams that already run agent infrastructure and need models that can autonomously execute repository-level developer pipelines.

Hardware requirements vary across the family. DeepReinforce notes a 9 billion-parameter model is small enough to run on powerful mobile or edge devices, while the 397 billion-parameter model requires substantial server infrastructure not available on consumer hardware. The MIT license and lack of regional restrictions are intended to facilitate self-hosted deployment and integration into private agentic workflows.

Ornith-1.0, its model card and evaluation results are available on Hugging Face and in the project’s documentation.