DeepReinforce AI Unveils Ornith-1.0: A Self-Improving Open-Source Model Family for Agentic Coding

What Changed

DeepReinforce AI has announced the release of Ornith-1.0, a new family of open-source models specifically engineered for agentic coding. This release includes several model variants: 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE. The models are post-trained on top of Gemma 4 and Qwen 3.5, and are designed to achieve state-of-the-art performance among open-source models of comparable size on various coding benchmarks. A key innovation highlighted is Ornith-1.0's self-improving training framework, which utilizes reinforcement learning (RL) to jointly optimize solution rollouts and the scaffolding that drives them.

The Ornith-1.0 models are MIT licensed, ensuring global accessibility without regional restrictions. The initial release focuses on the Ornith-1.0-9B model, which is optimized for efficient single-GPU deployment. DeepReinforce AI has also made GGUF versions available for the 9B and 35B models, facilitating deployment on platforms like Ollama and llama.cpp.

Technical Details

Ornith-1.0 is built upon a self-improving training framework that leverages reinforcement learning. This framework enables the models to learn not only to generate solution rollouts but also the underlying scaffold that guides these rollouts. By optimizing both the scaffold and the resulting solution, the model aims to discover more effective search trajectories and produce higher-quality code. The model family includes dense and Mixture-of-Experts (MoE) architectures, with sizes ranging from 9 billion to 397 billion parameters.

The Ornith-1.0-9B model, a dense 9B parameter model, requires approximately 19 GB of memory in bf16 precision, making it suitable for deployment on a single 80GB GPU. It is designed as a reasoning model, where the assistant's turn typically begins with a <think> … </think> block before the final answer. Serving recipes provided for vLLM and SGLang enable a reasoning parser to return the chain-of-thought in a separate reasoning_content field, and a tool-call parser to surface <tool_call> blocks as OpenAI-style tool_calls.

Deployment of Ornith-1.0-9B necessitates recent runtimes, specifically Transformers version 5.8.1 or newer, vLLM version 0.19.1 or newer, and SGLang version 0.5.9 or newer. Recommended sampling parameters include temperature=0.6, top_p=0.95, and top_k=20, with temperature=1.0 used for benchmark reproduction. The models support an OpenAI-compatible API endpoint, allowing integration with existing OpenAI clients and agent frameworks. This compatibility extends to streaming tokens and tool-calling capabilities, where Ornith-1.0-9B emits well-formed function calls parsed into the standard tool_calls field.

Benchmark Analysis

Ornith-1.0-9B demonstrates competitive performance across several agentic coding benchmarks. On Terminal-Bench 2.1 (Terminus-2), Ornith-1.0-9B achieved a score of 43.1, surpassing Qwen3.5-9B (21.3) and Gemma4-12B (21), and closely approaching Qwen3.5-35B (41.4) and Gemma4-31B (42.1). For Terminal-Bench 2.1 (Claude Code), Ornith-1.0-9B scored 40.6, outperforming Qwen3.5-9B (18.9) and nearing Qwen3.5-35B (38.9).

In SWE-bench evaluations, Ornith-1.0-9B recorded 69.4 on Verified, 42.9 on Pro, and 52 on Multilingual. These scores are higher than Qwen3.5-9B (53.2, 31.3, 39.7 respectively) and Gemma4-12B (44.2, 27.6, 32.5 respectively). On NL2Repo, Ornith-1.0-9B scored 27.2, exceeding Qwen3.5-9B (16.2), Qwen3.5-35B (20.5), Gemma4-12B (10.3), and Gemma4-31B (15.5). For Claw-eval Avg, Ornith-1.0-9B achieved 63.1, outperforming Qwen3.5-9B (53.2), Gemma4-12B (32.5), and Gemma4-31B (48.5), while being slightly below Qwen3.5-35B (65.4).

SWE Atlas benchmarks also show Ornith-1.0-9B's performance: 17.9 on QnA, 16.6 on RF, and 15.3 on TW. These results are notably higher than Qwen3.5-9B (9.2, 4.3, 4.4 respectively) and Qwen3.5-35B (13.2, 10.2, 9.8 respectively).

Developer Implications

The release of Ornith-1.0 provides developers with a new family of open-source models specialized for agentic coding. The MIT license allows for broad adoption and integration into commercial and non-commercial projects without restrictive licensing concerns. The availability of various model sizes, including the lightweight 9B version, facilitates deployment across different hardware configurations, from single 80GB GPUs to distributed setups using --tensor-parallel-size.

Ornith-1.0's OpenAI-compatible API endpoint simplifies integration into existing developer workflows and agent frameworks such as Hermes Agent, OpenClaw, and OpenHands. The model's explicit reasoning (<think> … </think>) and tool-calling capabilities can enhance the transparency and functionality of agentic applications. Developers can leverage the reasoning_content field to inspect the model's thought process, which can be valuable for debugging and understanding agent behavior. The support for GGUF builds further extends accessibility to platforms like Atomic.chat, Ollama, and llama.cpp, broadening the potential user base for local inference and experimentation.

For those working with coding CLIs, Ornith-1.0-9B can be configured as a provider, enabling terminal-based coding agents to interact with the model for tasks such as codebase understanding and automation. The model's optimization for terminal-based agents suggests its utility in environments where direct interaction with code and shell commands is paramount. The requirement for recent runtime versions (Transformers, vLLM, SGLang) implies that developers will need to maintain up-to-date environments to fully utilize Ornith-1.0's features and performance.

Bottom Line

DeepReinforce AI's Ornith-1.0 model family represents a contribution to the open-source landscape for agentic coding. Its self-improving RL framework for joint optimization of solution and scaffold generation offers a method for enhancing code quality and search trajectories. The models demonstrate competitive performance on established coding benchmarks, particularly in agentic coding tasks. With MIT licensing, OpenAI-compatible API endpoints, and support for various deployment runtimes and agent frameworks, Ornith-1.0 is positioned for integration into developer toolchains. The availability of GGUF versions and the focus on efficient single-GPU deployment for the 9B model further enhance its accessibility for a range of development and research applications.

Pneumetron