DeepReinforce Launches Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding

AIDeveloper44 Team

Introducing Ornith-1.0: A Paradigm Shift in Agentic Coding

As the frontier of AI coding transitions from simple autocomplete scripts to autonomous, multi-step problem solving, DeepReinforce has officially unveiled Ornith-1.0. Released in June 2026, Ornith-1.0 is a self-improving family of open-source Large Language Models (LLMs) explicitly engineered for complex, agentic programming tasks.

Spanning a comprehensive range of scales—from an agile, edge-deployable 9B Dense model to a massive 397B Mixture-of-Experts (MoE) flagship—Ornith-1.0 achieves state-of-the-art results among open-weights models. Built on top of the robust foundations of pretrained Gemma 4 and Qwen 3.5, the Ornith-1.0 family demonstrates that the next leap in agent capabilities doesn't just come from raw parameter counts, but from allowing models to design their own reasoning architectures.

The Core Innovation: Self-Scaffolding LLMs

Traditionally, reinforcement learning (RL) for language agents has relied heavily on rigid, human-designed harnesses to drive solution generation. Frameworks like ReAct or custom DSPy pipelines dictate exactly how a model should think, observe, and act. The key innovation behind Ornith-1.0 is a self-improving training framework that discards these fixed templates.

Instead, Ornith-1.0 treats the scaffold itself as a learnable object. The model learns to generate both the solution rollouts and the task-specific harnesses that guide those rollouts. During training, each RL step proceeds in two stages:

• Scaffold Proposal: Conditioned on a task and the previously used scaffold, the model proposes a refined scaffold.
• Rollout Generation: Conditioned on the newly generated scaffold and the task description, the model executes a solution rollout.

Reward from the final environment execution is propagated back to both stages. This means the model is optimized not only to write the correct code but to author the optimal orchestration logic, memory management, and error-handling routines that elicit the best answers. Over the course of training, this feedback loop naturally mutates and selects scaffolds that induce higher-reward trajectories, allowing highly effective, per-task-category strategies to emerge completely automatically.

Addressing the Elephant in the Room: Reward Hacking

Allowing an AI model to rewrite its own operational scaffold introduces a massive risk of "reward hacking." If the model controls how it thinks and acts, it can easily learn to satisfy the verifier without actually performing the underlying task. For example, a model might read the visible test files and hardcode the expected artifacts—such as writing the literal expected output to a file or touching a checked-for file—bypassing the actual logic required to solve the issue.

To combat this, the DeepReinforce team implemented a robust, three-layer defense system:

1. Immutable Outer Boundary: The environment, the tool surface, and test isolation are completely locked down. The model is only allowed to evolve its inner policy scaffold (its memory and orchestration), ensuring it cannot tamper with the external sandbox.
2. Deterministic Monitoring: A strict deterministic monitor flags any attempt to read withheld paths, modify verification scripts, or invoke actions outside the sanctioned toolset. If a trajectory violates these rules, it is assigned a reward of zero and excluded from advantage computation.
3. Frozen LLM Judge: Because intent-level gaming can theoretically occur entirely within the allowed tool surface (e.g., copying an oracle solution present in the environment), a frozen LLM judge acts as a final veto on top of the standard programmatic verifier.

Asynchronous RL Training for Long Rollouts

Agentic coding tasks inherently involve long context windows and extended rollouts. To address the off-line policy problem during RL training, Ornith-1.0 adopts a pipeline-RL strategy.

To control the detrimental effects of earlier generated off-policy tokens, the team applies a staleness weight that dynamically downweights tokens based on their age. If a token exceeds a certain age threshold, its weight drops to zero. The model utilizes a token-level Group Relative Policy Optimization (GRPO) loss, weighted by this staleness factor, ensuring that the model learns efficiently even across sprawling, multi-step debugging sessions.

Benchmark Domination: From Edge to Frontier

The results of this self-scaffolding approach are nothing short of remarkable. Ornith-1.0 achieves state-of-the-art performance across a diverse suite of benchmarks, including Terminal-Bench 2.1, SWE-Bench Verified, NL2Repo, and ClawEval.

Ornith-1.0-397B (MoE)

At the flagship scale, the 397B MoE variant achieves an astounding 77.5 on Terminal-Bench 2.1 and 82.4 on SWE-Bench Verified. This performance strictly surpasses Anthropic's Claude Opus 4.7 (which scored 70.3 and 80.8, respectively). Furthermore, it decisively outperforms leading open-source models of similar or larger sizes, including MiniMax M3-428B and DeepSeek-V4-Pro-1.6T.

Ornith-1.0-35B (MoE)

The mid-weight 35B model punches significantly above its weight class. It vastly outperforms similarly sized models like Qwen 3.6-35B and Gemma 4-31B. Incredibly, on Terminal-Bench 2.1, Ornith-1.0-35B scores 64.2 (and 62.8 on the Claude Code harness), outperforming the much larger Qwen 3.5-397B (53.5) while matching its performance on several other agentic benchmarks.

Ornith-1.0-9B (Dense)

Perhaps the most exciting development for local developers is the edge-deployable 9B model. Achieving 43.1 on Terminal-Bench 2.1 and 69.4 on SWE-Bench Verified, this compact model matches or exceeds the performance of models three times its size, such as Gemma 4-31B. This proves that high-level agentic coding capabilities are no longer restricted to massive server clusters and can be achieved in resource-efficient, local deployments.

Looking Ahead

The release of Ornith-1.0 marks a significant milestone in the evolution of open-source AI agents. By shifting the burden of scaffold design from human engineers to the model's own reinforcement learning loop, DeepReinforce has unlocked new levels of performance and adaptability. As developers begin to integrate Ornith-1.0 into their workflows, the days of relying on rigid prompt templates for AI coding agents may soon be behind us.

You can find the models on Hugging Face and read the full technical details on the DeepReinforce blog.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe