9B dense; 35B total / approximately 3B active MoE; 397B MoE
262,144 tokens natively (approximately 1M with validated YaRN scaling)
MIT
MIT-licensed open-weight release of the Ornith-1.5 family across 9B dense, 35B-A3B mixture-of-experts, and flagship 397B mixture-of-experts scales. The released checkpoints use Qwen3.5 dense and MoE conditional-generation architectures, with official configurations retaining Qwen3.5 vision encoders and image and video token support. Ornith-1.5 expands the self-scaffolding reinforcement-learning framework introduced with Ornith-1.0 into a broader self-improvement loop in which the model proposes new training tasks, generates or refines task-specific scaffolds and evaluation harnesses, produces solution rollouts, and jointly optimizes all three stages with GRPO. Generated tasks are rewarded according to validity and verifiability, proximity to the model's current capability frontier, and novelty relative to previously generated tasks, while harnesses are evaluated for task alignment, reward fidelity, and resistance to reward hacking. All three checkpoints are reasoning models that emit separate thinking traces and structured tool calls suitable for OpenAI-compatible APIs and agent frameworks. They provide 262,144-token native context windows and support tested YaRN scaling to approximately one million tokens, although the team cautions that static scaling can reduce quality on ordinary-length requests. The official Hugging Face collection provides BF16 checkpoints for all three scales, GGUF versions for all three, FP8 and NVFP4 variants for the two MoE models, and MLX variants for the 9B and 35B models. Self-reported results averaged over five runs include 86.1 on Terminal-Bench 2.1, 86.0 on SWE-bench Verified, and 56.0 on DeepSWE for Ornith-1.5-397B; 68.5, 79.0, and 22.0 respectively for Ornith-1.5-35B-A3B; and 47.0 on Terminal-Bench 2.1 plus 70.6 on SWE-bench Verified for Ornith-1.5-9B. The evaluation methodology includes network isolation, removal of Git history, and other anti-hacking measures on several software-engineering benchmarks.