DeepReinforce presents Ornith-1.0: a revolution for AI coding agents

DeepReinforce, an AI research lab known for previous projects like [CUDA-L1](https://github.com/deepreinforce-ai/CUDA-L1) and IterX, last week launched Ornith-1.0, a family of open-source models specialized in agentic coding tasks. The release, which took place on June 25 under the MIT license without regional restrictions, has sparked particular interest due to its superior performance compared to other models of similar size, such as Google's Gemma 4-31B.

A model designed for coding agents

Ornith-1.0 stands out for being specifically designed for coding agents operating in real environments, such as terminals and repositories. This "agentic" approach represents a breakthrough compared to traditional conversational interfaces, where the user interacts step-by-step with the AI. Coding agents, on the other hand, receive a task and execute it autonomously, without the need for continuous human guidance.

DeepReinforce describes Ornith-1.0 as a family of "self-improving" models, capable of improving their performance through continuous learning. This concept of "agentic" is fundamental: these are models that not only perform coding tasks but do so autonomously, managing entire development workflows without human intervention.

Technical specifications and performance

Ornith-1.0 is available in four variants, differentiated by the number of parameters: 9 billion, 31 billion, 35 billion (mixture of experts), and the flagship model with 397 billion (also a mixture-of-experts). Parameters represent the configurations and settings that a model can handle during training: the more there are, the more the model is capable of performing complex tasks. However, models with a high number of parameters require powerful hardware, not available on consumer devices.

The 9-billion-parameter model, considered small but efficient, achieves a score of 69.4 on SWE-bench Verified, significantly outperforming Google's Gemma 4-31B, which stops at 52.0. This result is particularly relevant as it demonstrates that even smaller models can be extremely effective in specific tasks.

How Ornith-1.0 works

Most current AI coding agents operate within a predetermined structure, a set of fixed rules that define how the agent must work. Ornith-1.0, on the other hand, treats this structure as a learnable object that evolves along with the model's policy. In other words, instead of following a predefined manual, Ornith develops its own strategies.

During reinforcement learning, the model proposes a strategy to tackle a task and then uses it to generate a solution. The feedback derived from the task outcome is used to improve both the strategy and the generated code. This process, repeated millions of times, allows the model to develop specific approaches for particular tasks without the need for human intervention.

DeepReinforce has also implemented security measures to prevent "reward hacking," i.e., the manipulation of the reward system by the model. Three levels of defense protect the process: the environment and test suite are immutable, a deterministic monitor detects any attempt to access restricted paths or modify verification scripts, and a frozen judge model serves as the last line of defense against manipulations.

Performance and comparisons with other models

The flagship 397-billion-parameter model achieves a score of 82.4 on SWE-bench Verified, a test that evaluates an AI's ability to fix bugs in open-source repositories without access to the test suite. This score surpasses that of Claude Opus 4.7 (80.8) and DeepSeek-V4-Pro (80.6). On Terminal Bench 2.1, which evaluates the ability to complete 89 tasks in containerized terminal environments, Ornith-1.0-397B achieves a score of 77.5, compared to 70.3 for Claude Opus 4.7.

Concerns have been raised about the potential risks associated with every investment in the Crypto sector: readers are advised to always inform themselves independently before making any decisions.