Ornith-1.5: From Self-Scaffolding to Self-Improvement
ornith.ai
August 19, 2026
11 min read
55/100
Summary
Ornith introduced Ornith-1.5, a family of 397B-parameter mixture-of-experts (MoE), 35B MoE, and 9B dense models trained through a self-improvement loop. The system generates progressively harder tasks, creates task-specific scaffolds containing instructions, tools, decomposition, and orchestration, then produces reinforcement-learning rollouts. Rewards optimize task generation, scaffold construction, and solutions jointly using validity, frontier difficulty, novelty, solution quality, and resistance to reward hacking. The target task success rate is 0.2, favoring problems difficult enough to expose capability gaps while still yielding successful training trajectories. Ornith reports that its 397B model scored 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, compared with 85.0 and 59.0 for Claude Opus 4.8. It scored 86.0 on SWE-bench Verified, 92.8 on GPQA Diamond, and 80.0 on MCP-Atlas. The 35B MoE model activates 3B parameters per token and scored 68.5 on Terminal-Bench 2.1 using Claude Code and 79.0 on SWE-bench Verified. The 9B model scored 47.0 and 70.6 on those benchmarks, respectively; Ornith says its quantized Mobile version can run on iPhone and Android devices. Ornith reports that its benchmark results are averaged over five independent runs.
Key Takeaways
- Ornith-1.5 trains models by jointly generating new tasks, building task-specific evaluation scaffolds, and optimizing solution rollouts with reinforcement learning.
- Ornith-1.5-397B scored 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE; Ornith compares these results with Claude Opus 4.8 scores of 85.0 and 59.0.
- The 35B MoE model activates 3B parameters per token and scored 68.5 on Terminal-Bench 2.1 with Claude Code and 79.0 on SWE-bench Verified.
- Ornith-1.5-9B scored 47.0 on Terminal-Bench 2.1 with Claude Code and 70.6 on SWE-bench Verified, and its quantized Mobile version is designed for iPhone and Android deployment.
- Task rewards favor valid, verifiable, novel tasks near the model's capability frontier, with a target rollout success rate of 0.2.
What the discussion said
The thread mostly treated Ornith-1.5 as a practical local-model release, not a breakthrough in self-improving AI. Readers focused on whether its 35B and 9B variants can deliver useful coding, tool use, and agent behavior on machines people actually own. Early hands-on reports were encouraging: the 35B-A3B looked competitive with Qwen 3.8 27B for web scraping while running faster and at less damaging quantization, and another tester found efficient reasoning and tool use. The open-weight release and the prospect of a compact model approaching much larger proprietary systems also generated real excitement. But the comparisons were far from settled. Several commenters argued that mixture-of-experts does not magically make local inference easy: on ordinary discrete-GPU setups, total VRAM still dominates, and dense models may buy more capability at a given memory limit. Benchmark figures shared in the thread suggest Ornith trails Qwen 3.8 sharply on several coding evaluations, especially DeepSWE, even if it wins one repository-oriented task. Readers also questioned whether the company’s self-improvement framing means altered model weights or merely a stronger agent harness, and one prior 9B user found the earlier Ornith underperformed its Qwen base despite published scores. The 397B model was viewed as effectively datacenter-scale hardware territory, limiting its local appeal.
Where opinion split
The sharpest dispute is whether the MoE 35B model is the better route to capable local AI. Supporters point to faster real-world inference, higher usable quantization, and strong practical results versus Qwen 27B; skeptics argue that VRAM capacity, not active parameters, determines what most consumer GPUs can run, while Qwen 3.8 retains a meaningful coding-quality lead.
Community Sentiment
Positives
- The 35B-A3B reportedly handled web-scraping work on par with Qwen 3.8 27B while running much faster at Q4 rather than forcing a quality-eroding Q8 setup.
- Early testing found leaner reasoning and competent tool use, suggesting the model can feel quicker in agent loops because it spends less time on unproductive deliberation.
- Open weights give local-AI users a model family they can test, quantize, and deploy without depending on a hosted provider.
- A 9B model posting competitive headline results against a far larger proprietary system has readers hopeful that capable personal agents are becoming genuinely small-model territory.
Concerns
- Shared benchmark results put Ornith behind Qwen 3.8 27B on major coding tests, with the DeepSWE gap large enough to undermine claims of coding parity.
- MoE does not remove the local deployment bottleneck for typical discrete GPUs: loading the full model still consumes VRAM, where a dense alternative may provide more intelligence per feasible setup.
- The 397B variant is too large for modest enthusiast hardware, making its practical use dependent on multi-node accelerator infrastructure rather than a home lab.
- The self-improvement pitch remains opaque because readers cannot tell whether it changed learned weights or simply wrapped an existing model in a better agent scaffold.
