開放模型
Ornith-1.5 Brings Self-Generated Tasks Into Reinforcement Learning; 397B Model Self-Reports 86.1 on Terminal-Bench 2.1
Ornith has released three open-weight models: a 9B model, a 35B MoE, and a 397B MoE. The training loop jointly optimizes tasks, agent scaffolds, and solution trajectories. The 397B model’s coding-agent performance approaches that of closed-source frontier models, but the figures have not yet been independently reproduced, and the model’s release documentation has a licensing gap.
Ornith has announced the Ornith-1.5 series, comprising a 9B dense model, a 35B MoE that activates approximately 3B parameters per token, and a 397B MoE. Compared with Ornith-1.0, which optimized only agent scaffolds and solution trajectories, the new release also uses reinforcement learning to generate the next batch of training tasks. Based on its existing solution history, the model first proposes more difficult problems, then constructs task-specific scaffolds containing tools, decomposition strategies, and verifiers, and finally generates solution trajectories.
A task’s reward is calculated by multiplying three factors: validity, frontier difficulty, and novelty. Validity checks whether the environment can execute the task and verify whether the result is correct. The difficulty function favors problems with a success rate near 0.2, while the novelty factor discourages repeatedly generating similar problems. This multiplicative design gives tasks a zero or very low reward if they fail on any one factor, although its practical effectiveness still depends on whether the verifiers can identify specification loopholes and reward hacking.
According to Ornith’s self-reported results, the 397B model scored 86.1 on Terminal-Bench 2.1, 86.0 on SWE-bench Verified, and 56.0 on DeepSWE. Each result was averaged across five runs; for some evaluations, network access was disabled and Git history was removed. The model supports a 256K context window, and its BF16 weights require approximately 800GB. Ornith recommends tensor-parallel deployment across eight H200 GPUs and also provides FP8, INT4, GGUF, and MLX versions.
Engineering teams should not treat leaderboard results as direct evidence of production readiness. These scores were primarily measured by the publisher, some evaluations used Claude as the judge, and complete reports on training data, costs, and independent reproduction are still unavailable. A third-party review also found that although the model card labels the release as MIT-licensed, the referenced LICENSE file still returned a 404 as of August 21. Until the full license text is provided, enterprises should not rely solely on a platform label when deciding whether to redistribute the model or use it commercially.