The Ornith team has released Ornith-1.5, a family of open-weight language models built around an unusual idea: the model generates its own training tasks, designs its own scaffolds, and learns from its own solution attempts in a continuously running loop. According to the team's own evaluations, the flagship Ornith-1.5-397B scores 86.1 on Terminal-Bench 2.1 (Terminus-2), putting it on par with Anthropic's Claude Opus 4.8 at 85.0 and ahead of every other open-source model of comparable size.
The release, published this week on the Ornith blog and on Hugging Face under an MIT license, is the latest salvo in the open-source AI race that has regularly produced results once thought impossible without a frontier lab's budget. For developers and enterprises tracking the latest AI model news, the significance is twofold: the benchmark parity with a leading closed model, and a training recipe that points at where open model development is heading next.
Three Sizes, One Recipe
Ornith-1.5 ships in three configurations: a 397B-parameter mixture-of-experts flagship, a 35B MoE that activates only 3B parameters per token, and a compact 9B dense model with a quantized "Mobile" variant designed to run directly on iPhone and Android devices. All three are available on Hugging Face alongside GGUF, MLX, and NVFP4 builds, and the model cards indicate the family is built on the Qwen3.5 MoE architecture.
The lineage matters here. Ornith-1.0, released earlier this year, popularized the "self-scaffolding" approach to agentic coding and became a quiet hit — its 9B GGUF build has logged more than five million downloads on Hugging Face. Ornith-1.5 extends that framework from optimizing scaffolds and rollouts into a full self-improvement pipeline.
The Self-Improvement Loop, Explained
In a conventional post-training pipeline, reinforcement learning runs against a fixed set of human-curated tasks. Ornith-1.5 instead has the model itself propose new tasks, generate a task-specific scaffold — the instructions, tools, and decomposition strategy used to attack the problem — and then produce solution rollouts that are scored by the scaffold it just built. Reward is propagated back through all three stages using GRPO, so the system simultaneously learns to write better tasks, better scaffolds, and better solutions.
The task-generation reward is the heart of the design. Each proposed task is scored on three multiplicative signals: validity (does the scaffold execute and evaluate correctly?), frontier difficulty (is the model's empirical success rate near a target of roughly 20 percent, keeping tasks challenging but learnable?), and novelty (is it sufficiently different from everything in a buffer of previously generated tasks?). Because difficulty is measured against the model's own current success rate, the curriculum automatically escalates as the model improves.
The team also reports explicit anti-hacking measures during evaluation: git history is stripped from repository images so models cannot recover prior solutions, and network access is disabled to prevent models from fetching external answers. All results are averaged over five independent runs.
The Numbers
On agentic coding, the flagship's self-reported results cluster at the top of the open-source field. On Terminal-Bench 2.1 run through the Terminus-2 harness, Ornith-1.5-397B's 86.1 edges out Claude Opus 4.8 (85.0), DeepSeek-V4-Flash-0731 (82.7), and GLM-5.2 (81). On the same benchmark run through the Claude Code harness, Ornith-1.5 posts 85.2 against Opus 4.8's 78.9. On SWE-bench Verified, the two are effectively tied at 86.0 and 85.8, with Kimi K3 slightly ahead at 86.2.
The gaps widen on harder agentic suites. On DeepSWE, Ornith-1.5-397B scores 56.0 — ahead of GLM-5.2's 46.2, though behind Opus 4.8 (59.0) and Kimi K3 (67.5). The most striking jump is generational: Ornith-1.0 scored just 8.0 on DeepSWE, meaning the new iteration delivers a sevenfold improvement on the same benchmark family.
The smaller models tell their own story. Ornith-1.5-35B-A3B, activating only 3B parameters per token, scores 68.5 on Terminal-Bench 2.1 (Claude Code) versus 43.4 for Google's Gemma 4-31B and 51.7 for Meta's Muse Glimmer-30B, and posts 79.0 on SWE-bench Verified. The 9B model reaches 47.0 on Terminal-Bench 2.1, 70.6 on SWE-bench Verified, and 86.4 on GPQA Diamond — numbers that put a phone-class model within striking distance of desktop-class rivals.
Caveats and Context
These are developer-run benchmarks, not independently audited results, and the comparison models were evaluated by the Ornith team under its own harness settings. Rival labs also tune for different trade-offs; Kimi K3's lead on DeepSWE suggests the frontier of agentic software work is still contested. But the release's full-table transparency — five-run averages, published harness configurations, disclosed safeguards — makes independent reproduction unusually straightforward.
The community response has been substantial. The release announcement drew well over a hundred points on Hacker News within hours, with discussion focused less on the benchmark claims than on the self-improvement methodology, which several commenters described as one of the more credible open implementations of recursive-style task generation.
For the open-source ecosystem, Ornith-1.5 lands at a moment when open models from China's labs and Western independent teams have been closing the gap with closed frontiers on coding and agentic tasks. An MIT-licensed family that matches a leading closed model on Terminal-Bench — and ships a phone-ready 9B variant — narrows that gap further, and does it with a training method that anyone can inspect, reproduce, and extend.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
