DeepReinforce Launches Ornith-1.0: Open-Source AI Coding Models That Learn Their Own Scaffolds

DeepReinforce has officially launched Ornith-1.0, a new family of open-source AI models specifically engineered for agentic coding. This release is notable for its innovative approach where the models learn to construct their own operational scaffolds during reinforcement learning, departing from the conventional method of using pre-defined, human-engineered harnesses. The Ornith-1.0 family, built upon Gemma 4 and Qwen 3.5, offers a range of models, including a flagship 397B mixture-of-experts (MoE) variant, all released under the permissive MIT license.

Key facts

Feature	Detail
Developer	DeepReinforce
Model Family	Ornith-1.0
Core Innovation	Models learn to write their own RL scaffolds
License	MIT License (all weights)

The Ornith-1.0 models are designed as reasoning engines tuned for coding agents. The lineup includes four sizes: 9B Dense, 31B Dense, 35B MoE, and 397B MoE. The 35B MoE model activates approximately 3 billion parameters per token, balancing performance and efficiency. DeepReinforce has also published FP8 and GGUF builds, facilitating faster local deployment and accessibility for developers with varying hardware capabilities. This makes the technology more approachable for Indian startups and developers keen on leveraging advanced AI for coding tasks without extensive infrastructure.

A Differentiated Approach to AI Coding

Unlike most coding agents that pair a model with a fixed, human-designed scaffold (also known as a harness), Ornith-1.0 treats the scaffold as a learnable component. During the reinforcement learning process, the scaffold co-evolves with the model’s policy. This two-stage learning process involves the model reading the task and its current scaffold, proposing a refined scaffold, and then using that scaffold to generate a solution. The reward from the solution generation feeds back into both stages, optimizing the model to author orchestration logic alongside generating code. This mechanism allows for the emergence of per-task strategies without the need for manual harness design.

Performance Benchmarks

DeepReinforce reports competitive performance for Ornith-1.0 across several agentic coding benchmarks. The flagship Ornith-1.0-397B model achieved a score of 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified. On SWE-Bench Verified, this places it just behind Claude Opus 4.8 (87.6) among listed models. While its performance on Terminal-Bench 2.1 is more varied, beating Claude Opus 4.7 (70.3) but trailing Claude Opus 4.8 (85) and GLM-5.2-744B (81.0), DeepReinforce emphasizes that these are state-of-the-art results for open models of comparable size. The smaller 35B model also demonstrates efficiency, scoring 64.2 on Terminal-Bench 2.1, surpassing Qwen 3.5-397B’s 53.5. The 9B model, suitable for edge or single-GPU setups, achieved 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified.

Practical Applications for Developers

These models are designed for terminal-native coding agents and repository-scale work, making them ideal for tasks such as multi-file refactoring, bug localization, and test-driven patches. For Indian developers and startups, the efficiency of the 9B model allows for local deployment to triage failing test suites, while the 397B model can be self-hosted by platform teams for internal coding agents, offering maximum accuracy on complex, multi-step tasks. The models also provide an OpenAI-compatible endpoint, ensuring compatibility with standard agent frameworks without requiring code changes.

Serving the Ornith-1.0 models is streamlined, with support for vLLM, SGLang, and Transformers. The 9B model, for instance, requires approximately 19GB in bf16 and can run on a single 80GB GPU. The availability of these models under the MIT license encourages widespread adoption and innovation within the Indian AI and technology ecosystem.

Source: MarkTechPost (https://www.marktechpost.com/2026/06/25/deepreinforce-releases-ornith-1-0-an-open-source-coding-model-family-that-learns-its-own-rl-scaffolds/)