Ornith-1.5 Release: End-to-End Self-Improvement for Reasoning and Coding

Ornith-1.5 is a series of foundation models designed for high-performance reasoning, agentic, and coding tasks. It evolves the self-scaffolding framework of Ornith-1.0 into a complete self-improvement loop where the model autonomously proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning.

Model Performance and Scaling

Ornith-1.5 is released in three scales, demonstrating competitive performance against both proprietary and open-source models across various benchmarks.

Ornith-1.5-397B (MoE)

The flagship 397B Mixture-of-Experts (MoE) model performs on par with Claude Opus 4.8 in several key benchmarks. It scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, outperforming open-source peers like GLM-5.2 (82.7 and 46.2) and DeepSeek-V4-Flash-0731 (82.7 and 54.4).

Ornith-1.5-35B (MoE)

The 35B MoE model activates only 3B parameters per token but significantly outperforms larger dense models. On Terminal-Bench 2.1, it scores 68.5, compared to 43.4 for Gemma 4-31B and 51.7 for Muse Glimmer-30B. It also achieves 79.0 on SWE-Bench Verified, surpassing both Gemma 4-31B (52.0) and Muse Glimmer-30B (76.0).

Ornith-1.5-9B (Dense)

Designed for edge deployment on iPhone and Android devices, the 9B dense model matches or exceeds the performance of much larger models. It achieves 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-Bench Verified, outperforming Gemma 4-31B and Qwen 3.6-35B in several categories.

The Self-Improvement Loop

Ornith-1.5 replaces human-curated tasks and manual harnesses with an autonomous cycle of task generation, scaffold construction, and solution production. This closed loop allows the model to continuously expand its own curriculum and adapt its problem-solving strategies.

Three-Stage Training Cycle

Task Proposal: The system proposes progressively harder tasks based on the model's previous history and the target environment, specifically targeting capability gaps.
Scaffold Generation: For each task, the model generates a task-specific scaffold, which includes the instructions, tools, decomposition strategy, and orchestration required to solve the problem.
Solution Rollout: The policy produces a solution rollout conditioned on the task and scaffold.

Reward from the rollout is propagated back through all three stages, optimizing the model's ability to generate useful tasks and effective scaffolds via Group Relative Policy Optimization (GRPO).

Reward Mechanisms

To ensure the self-improvement loop produces high-quality training data, Ornith-1.5 utilizes specific reward signals for tasks, harnesses, and rollouts.

Task Reward

Task rewards are calculated based on a multiplicative formula of three signals:

Validity: Ensures the task is coherent, solvable, and the scaffold executes correctly. This acts as a hard gate to prevent malformed tasks from being rewarded.
Frontier Difficulty: Measures if the task is challenging but learnable. The system targets a success rate ($π_{target}$) of 0.2, pushing the generator toward harder problems as the model improves.
Novelty: Reduces redundancy by rewarding tasks that are diverse relative to a buffer of previously generated tasks.

Harness and Rollout Rewards

Harness Reward: The harness is rewarded based on its alignment with the task specification, its ability to track true solution quality, and its resistance to reward hacking.
Rollout Reward: Each rollout is scored by the generated harness, using binary pass/fail for verifiable tasks or a combination of correctness, efficiency, and constraint satisfaction for complex environments.

Community Insights and Observations

User discussions highlight both the practical utility and the technical skepticism surrounding the release:

Local Deployment: Users have noted the efficiency of the 35B-A3B model, with one reporting it as being on par with Qwen 3.8 27B while maintaining higher speeds at higher quantization levels.
Model Identity: Some users observed that the model occasionally claims to be Claude in identity prompts.
Technical Skepticism: Some community members questioned whether the "self-improvement" occurs at the weight level or is simply an agentic wrapper, while others expressed skepticism regarding the claims of Recursive Self-Improvement (RSI) as a marketing tactic for funding rounds.