DeepReinforce has released Ornith-1.0, an open-source family of agentic coding models, with a 9B dense model that the company says can run on a single GPU and score 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified.
DeepReinforce has released Ornith-1.0, an open-source family of agentic coding models aimed at software engineering tasks.
According to DeepReinforce’s release page, the family includes an Ornith-1.0-9B model that the company describes as a dense, approximately 9-billion-parameter model designed for agentic coding workflows. The official Hugging Face model card for deepreinforce-ai/Ornith-1.0-9B also presents the 9B model as suitable for single-GPU use and includes serving examples with a 262,144-token context setting.
DeepReinforce reports that Ornith-1.0-9B scores 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified. The same figures appear on the official Hugging Face model card, which lists benchmark rows for Terminal-Bench 2.1 and SWE-Bench Verified.
Tech Times separately reported the same benchmark numbers in its coverage of the release, describing Ornith-1.0 as an open-source coding model family spanning 9B through 397B variants. GIGAZINE also covered the release and noted DeepReinforce’s positioning of the smallest 9B model for edge-device use, while citing the company’s comparison with larger models such as Gemma 4-31B.
These benchmark claims should be read as reported results from the model publisher and its official materials unless independently reproduced. The available source excerpts do not provide a full methodology for the benchmark runs beyond the named evaluation suites and scores.
DeepReinforce describes Ornith-1.0 as using “self-scaffolding” reinforcement learning for agentic coding. In the company’s framing, the model is trained not only to produce code, but also to construct intermediate scaffolds for solving software tasks.
Tech Times characterized this approach as a model that “writes its own training scaffold” during reinforcement learning, echoing DeepReinforce’s description. The practical claim is that Ornith-1.0 is intended for coding-agent behavior, where a model plans, edits, tests, and iterates across a task rather than only generating isolated code snippets.
The Ornith-1.0-9B model is available through Hugging Face under the deepreinforce-ai/Ornith-1.0-9B repository. Tech Times reported that the broader Ornith-1.0 family is released under the MIT license on Hugging Face.
The official model card describes the 9B version as a dense model and provides examples suggesting long-context serving configurations. For developers, the significance is that DeepReinforce is offering a smaller member of the family alongside larger variants, rather than limiting the release to frontier-scale models.
The confirmed facts from the cited sources are that DeepReinforce announced Ornith-1.0, published an official Hugging Face model card for the 9B model, and reported benchmark scores of 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified. Third-party outlets Tech Times and GIGAZINE corroborated the release and repeated key details about the model family and its positioning.
What remains less clear from the available source material is how broadly the reported performance generalizes across real-world repositories, hardware configurations, and developer workflows. Independent benchmark replication and hands-on evaluations would be needed to assess how Ornith-1.0-9B performs outside the publisher’s reported results.
DeepReinforce has released Ornith 1.0, an open source family of agentic coding models aimed at software engineering tasks.
According to DeepReinforce’s release page, the family includes an Ornith 1.0 9B model that the company describes as a dense, approximately 9 billion parameter model designed for agentic coding workflows.
The official Hugging Face model card for deepreinforce ai/Ornith 1.0 9B also presents the 9B model as suitable for single GPU use and includes serving examples with a 262,144 token context setting.
Continue reading
