Build and debug code with self-improving agentic models

Ornith is a family of self-improving open-source models for agentic tasks, covering two generations. Ornith-1.5 (the latest) extends self-scaffolding into a complete end-to-end self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning. Ornith-1.0, the first release, jointly optimizes the scaffold and the resulting solution rollouts. Both generations are MIT licensed, globally accessible, and free from regional limitations, with model weights published on separate HuggingFace collections for each generation.

Use it when you need an open-weight model tuned for agentic coding, reasoning, and tool-use tasks - published benchmark results cover coding (Terminal-Bench 2.1, SWE-bench Verified/Pro/Multilingual, NL2Repo), reasoning (HLE, GPQA Diamond), and agentic evaluations (MCP-Atlas, Toolathlon-Verified, WideSearch, BrowseComp, ClawEval). Three sizes are benchmarked - 397B, 35B-A3B, and 9B - so the model can be matched to the available compute.

On the published benchmarks, Ornith-1.5-397B scores 86.1 on Terminal-Bench 2.1 (Terminus-2), 86 on SWE-bench Verified, 65.1 on SWE-bench Pro, 92.8 on GPQA Diamond, 80 on MCP-Atlas, and 86.6 on BrowseComp - consistently ahead of Ornith-1.0-397B on the same benchmarks (77.5, 82.4, 62.2, 88.1, 76.4, and 79.7 respectively) and competitive with DeepSeek-V4-Flash-0731, GLM-5.2, Claude Opus 4.8, and Kimi K3 on the same table. The generation-over-generation gap is largest on DeepSWE (56 vs 8) and Frontier-Bench v0.1 (13.5 vs 2.7) at 397B, where Ornith-1.0 trailed most of the comparison models and Ornith-1.5 moved ahead of several of them.

The smaller sizes follow the same pattern. At 35B-A3B, Ornith-1.5 scores 67.8 on Terminal-Bench 2.1 (Terminus-2) and 79 on SWE-bench Verified versus Ornith-1.0's 64.2 and 75.6, and versus Qwen3.6-35B-A3B's 52.5 and 73.4; on GPQA Diamond it reaches 89.2 and on MCP-Atlas 70.2. At 9B, Ornith-1.5 scores 46.2 on Terminal-Bench 2.1 (Terminus-2) and 70.6 on SWE-bench Verified, against Qwen3.5-9B's 21.3 and 53.2 at the same size, with GPQA Diamond at 86.4 and MCP-Atlas at 54.2. Every size and category reported shows Ornith-1.5 ahead of Ornith-1.0 at the same parameter count. On the SWE-bench Multilingual, NL2Repo, and SWE Atlas - QnA rows the same ordering holds across all three sizes, and on HLE (both with and without tools) Ornith-1.5 also leads its predecessor at every size, though Claude Opus 4.8 still scores higher on the 397B GPQA Diamond and HLE rows.

Teams evaluating or deploying open-weight models for coding agents, tool-use, and agentic reasoning workloads who want published, size-comparable benchmark numbers against both open (Qwen, Gemma) and closed (Claude Opus, Kimi K3) models before choosing a weight class.