Ornith-1.5 models invent their own tasks to self-improve

Ornith-1.5 is a new family of open-source models built to train and improve itself without human supervision. Built on top of Qwen3.5 and Gemma 4 via continued pretraining, mid-training, and post-training, the lineup arrives in three sizes: a massive 397B MoE, an ultra-efficient 35B MoE, and a compact 9B Dense model.

Performance-wise, the flagship Ornith-1.5-397B goes toe-to-toe with frontier closed systems, matching Claude Opus 4.8 across grueling coding benchmarks. Meanwhile, the lightweight Ornith-1.5-9B model outperforms larger 30B-class models, and its quantized Ornith-1.5-9B-Mobile build can run entirely on-device across iPhone and Android hardware.

Inventing the curriculum on the fly

Standard reinforcement learning usually hits a ceiling once an AI exhausts human-crafted benchmark sets. Ornith-1.5 tackles this by turning task design into a closed self-improvement loop across three stages: task generation, scaffold construction, and solution rollouts.

First, the model inspects its environment and previous performance history to propose fresh tasks that lie right at its capability frontier. A multi-signal reward system scores these tasks on validity, difficulty, and novelty. To prevent the model from generating trivial questions or impossible puzzles, the developers target a 20% empirical success rate, pushing the model right to the edge of what it can barely solve.

Second, the model builds its own task-specific scaffolding, creating the instructions, tool interfaces, decomposition tactics, and automated test harnesses to evaluate candidate answers. Finally, the model attempts to solve the problem inside that harness.

Rewards from rollout performance propagate all the way back across the entire chain. Using GRPO, the system updates the question proposer, the test harness builder, and the problem solver in a single joint loop.

Benchmark results across all three scales

Across terminal navigation and software engineering benchmarks, Ornith-1.5 posted numbers that rival the best in the industry.

The 397B flagship scored 86.1 on Terminal-Bench 2.1 under Terminus-2 and 56.0 on DeepSWE, alongside 86.0 on SWE-bench Verified. That essentially ties Claude Opus 4.8 while comfortably beating open-weight peers like GLM-5.2 and DeepSeek-V4-Flash-0731.

The 35B MoE model activates just 3B parameters per token, but it hit 68.5 on Terminal-Bench 2.1 under Claude Code and 79.0 on SWE-bench Verified. That completely outpaces Qwen 3.6-35B as well as dense competitors like Gemma 4-31B and Muse Glimmer-30B.

Even the 9B dense model scored 47.0 on Terminal-Bench 2.1 and 70.6 on SWE-bench Verified, outpacing older open models twice its size.

Why it matters

We think Ornith-1.5 represents a major proof-of-concept for synthetic data generation: instead of mining human code repos or relying on static instruction sets, models can construct their own synthetic curricula and test environments to climb benchmark leaderboards.

For developers building local developer tools or on-device mobile agents, the quantized 9B-Mobile version is an immediate must-try. The catch, however, lies in production stability. Relying on an AI to write both the test questions and the grading harness carries the risk of reward hacking, meaning engineers deploying Ornith in mission-critical workflows will still want robust human guardrails.