Ornith-1.5 Hits SOTA 9B/35B and Matches Claude Opus 4.8 at 397B

DeepReinforce has released Ornith-1.5 under the MIT license, presenting a new foundation‑model framework that moves toward end‑to‑end self‑improvement by letting the model generate its own tasks, construct evaluation scaffolds, and produce solution rollouts.

The family includes three sizes: a 9B dense model, a 35B mixture‑of‑experts (MoE) model, and a flagship 397B MoE model. The 9B and 35B variants achieve state‑of‑the‑art results within their parameter class, while the 397B version reaches performance on par with Claude Opus 4.8.

On the Terminal‑Bench 2.1 suite the 397B model scores 86.1, close to Claude Opus 4.8’s 85.0, and on DeepSWE it attains 56.0 versus 59.0 for Opus. It also surpasses open‑source peers such as GLM‑5.2 (82.7 / 46.2) and DeepSeek‑V4‑Flash‑0731 (82.7 / 54.4).

The 35B MoE model activates only 3 B parameters per token and consistently outperforms the same‑size Qwen 3.6‑35B on coding and agent benchmarks, demonstrating that selective activation can retain strong performance.

The smallest 9B model uses a dense architecture; after quantization it can run directly on standard smartphones (iPhone and Android). Despite its size, it scores 47.0 on Terminal‑Bench 2.1 and 70.6 on SWE‑Bench Verified, exceeding larger models such as Gemma 4‑31B and Qwen 3.6‑35B.

The self‑improvement loop consists of three tightly coupled stages. First, the model proposes new, higher‑difficulty tasks by analysing its current environment and past solutions, deliberately targeting its own capability gaps. Second, it builds a dedicated scaffold for each task, specifying commands, external tools, problem‑decomposition strategies, and execution orchestration. Third, guided by the task and scaffold, the model generates a full solution rollout, converting theory into concrete execution steps. Rewards obtained during rollout are back‑propagated to the earlier stages, enabling the model to learn not only better answers but also how to formulate valuable training tasks and construct more efficient scaffolds.

By integrating task generation, scaffold construction, and solution rollout into a single optimization loop, Ornith‑1.5 reshapes the traditional static‑distribution training paradigm and demonstrates that large language models can autonomously evolve their own curricula.