On August 18, Ornith AI released the open-source Ornith-1.5 family in three scales at once: a 397B MoE, a 35B MoE (activating roughly 3B parameters per token), and a 9B dense model — the last one shipping a quantized Mobile variant that deploys directly on iPhone and Android devices. The theme of this generation is not parameter count; it is extending the self-scaffolding framework introduced in Ornith-1.0 back in June into an end-to-end self-improvement loop.
The self-improvement loop: the model writes its own tasks and builds its own exam
The official blog's description of the loop is fairly aggressive: given an environment or codebase, the model proposes progressively harder new tasks, deliberately targeting types it has not yet solved; it then generates or refines a task-specific scaffold — instructions, tools, decomposition strategy, and orchestration — for each task; finally, conditioned on the task and scaffold, the policy produces a solution rollout. Reward from the rollout propagates back across all three stages, jointly optimized with GRPO. In other words, the system does not just learn to solve better problems — it simultaneously learns to pose more training-worthy problems and build more effective exam rooms.
The task-side reward is the product of three signals: validity (a task must be verifiable; failing validation zeroes the reward outright), frontier difficulty (estimated from the model's own rollout success rate, with the target set at 0.2 — once a task is reliably solved, its reward decays, forcing the proposer to keep escalating difficulty), and novelty (deduplicated against a buffer of previously generated tasks to prevent re-grinding the same type). The harness-side reward is likewise a product of three terms: task alignment, reward fidelity, and resistance to reward hacking. Together this turns "curriculum" from a human-curated artifact into something the model produces and consumes itself.
The scorecard: 397B closes in on the closed frontier in agentic coding
In the published evaluations, the 397B flagship is strongest on the agentic coding line: 86.1 on Terminal-Bench 2.1 (Terminus-2), above Claude Opus 4.8's 85.0; 56.0 on DeepSWE, close to Opus 4.8's 59.0; and 86 on SWE-bench Verified, in the same tier as Opus 4.8's 85.8 and Kimi K3's 86.2. Against open-source rivals of similar scale, it leads GLM-5.2 (Terminal-Bench 82.7, DeepSWE 46.2) and DeepSeek-V4-Flash-0731 (82.7 and 54.4).
The 35B tier deserves more attention: 67.8 on Terminal-Bench 2.1 (Terminus-2), leading Qwen3.6-35B (52.5), Gemma-4-31B (42.1), and Meta's Muse Glimmer-30B (51.7) by more than 15 points; SWE-bench Verified 79.0 versus 73.4, 52.0, and 76.0. On DeepSWE the 35B scores 22, where both the previous Ornith-1.0-35B and Qwen3.6-35B scored 0. The 9B tier posts 70.6 on SWE-bench Verified and 47.0 on Terminal-Bench, clearly beating the 31B Gemma 4 and closing in on the 35B Qwen3.6 — and given it is only 9B and fits on a phone after quantization, edge teams should take a serious look.
The generational jump of the 397B itself is steep too: DeepSWE climbed from 8 in version 1.0 to 56, and Frontier-Bench v0.1 from 2.7 to 13.5.
Boundaries and judgment
To be clear, "matching closed models" holds only on the agentic coding line. On Frontier-Bench v0.1, the 397B's 13.5 still trails Opus 4.8's 21.1 and Kimi K3's 23 by a wide margin; GPQA Diamond at 92.8 versus Opus 4.8's 93.6 remains behind. The open-source camp has still not crossed the closed frontier on the hardest tasks — and every one of these numbers is currently vendor-reported; community reproduction has not started.
But for the industry, Ornith-1.5's value may lie less in the scores than in the reproducibility of the method: it has written the full recipe of "synthetic tasks + self-graded harness + GRPO" — including the multiplicative reward design, the hard validity gate, and anti-cheating evaluation details (stripping git history and disabling networking during SWE-bench runs to block retrieval of existing answers) — into a public release blog post. As marginal returns on generic data thin out, the next round of capability growth will come either from bigger clusters or from models manufacturing their own training signal; Ornith has cast a 397B open-weight vote for the second route. Whether the self-improvement loop hits a "fooling itself" reward hole at higher difficulty remains to be seen — their three-signal product is the first line of defense, and the real answer will arrive when the community takes the weights and reproduces (source: ornith.ai/ornith_1_5.html).