AI BriefingOrnith-1.5: From Self-Scaffolding to Self-Improvement
Key point
Ornith-1.5 achieves open-source SOTA through self-improvement via self-generated tasks and reinforcement learning
Details
Ornith-1.5 is a foundation model featuring an end-to-end self-improvement loop where the model proposes new tasks, generates scaffolds, and creates solution rollouts for reinforcement learning.
It is available in three sizes: 397B MoE, 35B MoE, and 9B Dense, demonstrating top-tier performance among open-source models of comparable scale on reasoning, agentic, and coding tasks.
- 397B model: Scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, delivering performance on par with Claude Opus 4.8.
- 35B model: Outperforms comparable models including Qwen 3.6-35B, showing a significant gap on agentic coding benchmarks despite activating only 3B parameters per token.
- 9B model: Deployable on edge devices such as iPhone and Android, outperforming larger models like Gemma 4-31B.
Instead of relying on fixed human-curated tasks, the model autonomously generates an evolving curriculum that pushes beyond its current capability limits, fostering sustainable capability improvement.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
DeepSeek Releases DSec: Elastic Sandbox Infrastructure for Large-Scale Agentic Training
Trajectory Introduces Density-Aware Training to Optimize AI Cost-Per-TaskStanford Researchers Propose EXPO-FT for Universal Post-Training in Robotics