DeepReinforce Unveils Ornith-1.0: Self-Improving AI Models Redefine Agentic Coding

A New Paradigm in Agentic Coding

DeepReinforce has introduced Ornith-1.0, a new family of open-source, self-improving models designed specifically for agentic coding and complex software engineering workflows. Released in June 2026, the models are built to operate as proper coding agents rather than mere chat assistants, capable of integrating with MCP servers, utilising tools, and hooks, and functioning in iterative loops. The name "Ornith" itself, derived from the ancient Greek word for bird, symbolises the model's ability to "build its own nest before it flies," reflecting its unique self-improvement capabilities. The Ornith-1.0 family is available in several sizes, including a 9B Dense model, a 31B Dense variant, a 35B Mixture-of-Experts (MoE) model, and a flagship 397B MoE model. These models are post-trained on robust base architectures, specifically Gemma 4 and Qwen 3.5. Crucially, all Ornith-1.0 models are released under the MIT license, making them fully open, free for commercial use, and unburdened by regional limitations, a factor noted to remove legal friction for adoption by commercial teams.

The Self-Scaffolding Breakthrough

The core innovation underpinning Ornith-1.0 is its self-improving training framework, a mechanism that structurally differentiates it from most existing agentic coding systems. Unlike conventional models that rely on fixed, human-designed harnesses to guide solution generation, Ornith-1.0 learns to generate both solution rollouts and the task-specific harnesses that guide those rollouts. This "self-scaffolding" reinforcement learning approach, employing a technique like GRPO, allows the model to jointly optimise both the scaffold and the resulting solution. By doing so, it can discover better search trajectories, organise its own thinking process, and produce higher-quality solutions. This design also incorporates a default reasoning mode, generating "think blocks" internally before delivering final answers. To mitigate potential reward-hacking risks inherent in such self-improvement, DeepReinforce has implemented a three-layer defense system comprising locked boundaries, deterministic monitoring, and a frozen judge model.

Performance and Accessibility

Ornith-1.0 models have demonstrated state-of-the-art performance among open-source models of comparable size across a range of agentic coding benchmarks, including Terminal-Bench, SWE-Bench, NL2Repo, OpenClaw, and ClawEval. The flagship Ornith-1.0-397B model achieved scores of 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified. This performance is highly competitive, matching Claude Opus 4.7 (which scored 70.3 on TB-2.1 and 80.8 on SWE-Bench Verified) and surpassing other leading open-source models such as MiniMax M3 and DeepSeek-V4-Pro on these benchmarks. The Ornith-1.0-35B MoE model also showed significant gains, outperforming similarly sized models like Qwen 3.5-35B, Qwen 3.6-35B, and Gemma 31B. Notably, it even surpassed the much larger Qwen 3.5-397B on Terminal-Bench 2.1 (64.4 vs. 53.5). For broader accessibility, the compact 9B Dense model is designed for edge device deployment and can run locally on commodity hardware, including single GPUs. It achieves strong results, scoring 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified, matching or exceeding the performance of larger models like Gemma 4-31B and Qwen 3.6 35B. Local tests using Ollama demonstrated that the 9B model offers similar accuracy to Qwen 3.5 9B while being approximately three times cheaper, and up to twenty times cheaper on some tasks. Deployment is flexible, supporting vLLM, SGLang, Transformers, Docker, and compatible quantized local applications, with OpenAI-compatible endpoints for tool calling. This release marks a significant stride for open-source AI, presenting a powerful, commercially viable alternative that could rapidly accelerate the development of sophisticated autonomous coding agents.