Ornith 1.5 presents full self-improvement loop and matches cloud opus 4.8

Researchers at Ornith released version 1.5 of their model family today, extending the self-scaffolding framework introduced in version 1.0 into a complete closed loop. The model now proposes new tasks, builds dedicated scaffolding—instructions, tools, decomposition strategies, and orchestration—and generates rollouts from which it learns via reinforcement learning. Each iteration creates fresh learning experiences that push performance further.

The flagship, Ornith-1.5-397B (a mixture-of-experts architecture), scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, placing it head-to-head with Claude Opus 4.8 (85.0 and 59.0 respectively). It also outperforms open counterparts such as GLM-5.2 (82.7 and 46.2) and DeepSeek-V4-Flash-0731 (82.7 and 54.4). The technical report states the model was designed for strong general intelligence on reasoning, agentic and code tasks, and the numbers indeed show broad benchmark coverage, but the report is a manufacturer technical document, not peer-reviewed, and the comparison to Claude Opus 4.8 relies on only two benchmarks.

The mid-size Ornith-1.5-35B (also MoE) activates just 3 billion parameters per token and still surpasses Qwen 3.6-35B across all code and agentic benchmarks, as well as dense models Gemma 4-31B and Muse Glimmer-30B from Meta. On Terminal-Bench 2.1 it reaches 68.5 versus 43.4 for Gemma and 51.7 for Muse; on SWE-Bench Verified it scores 79.0 versus 52.0 and 76.0 respectively. The compact Ornith-1.5-9B attains 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-Bench Verified, results that match or exceed much larger models, and the quantized 9B-Mobile variant can be deployed directly on iPhone and Android.

Each training cycle follows three stages. First, the system receives an environment or codebase, high-level task instructions, and access to the prior solution history, then proposes harder tasks that expose capability gaps. Second, the model generates or refines dedicated scaffolding for each task. Third, given the task and scaffolding, the policy produces a solution. The reward from the solution is back-propagated through all three stages, teaching the system not only to produce better solutions but also to create more useful training tasks and more effective scaffolding. In the closed loop, a stronger policy enables harder tasks, evolving scaffolding reveals better abilities, and higher-quality rollouts provide stronger learning signals, all without reliance on a static training distribution or manually designed agents.

The technical report omits data volume, training cost, duration, and hardware details, making independent efficiency assessment difficult. It also does not publish results on broader benchmarks such as MMLU or GPQA that would allow comparison beyond code and agentic tasks. The models are available on Hugging Face with open weights, a significant step for reproducibility, but the training code and self-improvement pipeline have not yet been released. In short, Ornith 1.5 demonstrates that a self-improvement loop can operate at a level approaching leading closed models, at least in the code and agentic bottleneck, and the open question now is whether the approach will scale to other domains without sacrificing stability.