The Arrival of Self-Evolving Coding AI “Ornith-1.0”!

What Happened? Overview of the News

• Introduction of the Self-Improving Open Source Model “Ornith-1.0”: A suite of models specialized for agent-based coding tasks, built on Gemma 4 and Qwen 3.5. It features four variations ranging from 9B to 397B MoE.
• Implementation of “Self-Scaffolding”: This innovative framework allows the model to generate and optimize its own task-specific “scaffolding,” moving away from traditional, human-designed fixed verification and execution processes.
• Performance That Surpasses Claude Opus 4.7: The flagship 397B model achieved scores of 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1. Even the lightweight 9B model performs comparably to the larger Gemma 4-31B.

Why Is This Important? Key Points to Note

• Automatic Evolution of the “Inference Process”: It’s groundbreaking that AI not only provides answers but also co-evolves the control logic (scaffolding) that dictates “how to reach those answers” through reinforcement learning (RL).
• Impenetrable Defense Against Reward Hacking: To prevent the model from creating convenient “scaffolding” and cheating, it employs a three-layer defense mechanism consisting of immutable environmental boundaries, deterministic monitoring, and frozen LLM judges.
• Scalability from Edge to Frontier: The 35B model even outperforms the significantly larger Qwen 3.5 in some benchmarks, showcasing exceptional size efficiency.

🦈 Shark’s Perspective (Curator’s Insight)

Ornith-1.0 is no ordinary coding AI! The idea of teaching the model how to scaffold itself pushes the boundaries of agent-based AI. Previously, humans dictated coding and debugging procedures, but now Ornith says, “For this task, I’ll devise my own memory management and error handling system!” It’s crafting strategies on its own, and that’s a game-changer! Especially impressive is the performance of the 9B model on edge devices. We’ve officially entered an era where lightning-fast local environments are powered by a brain on par with Claude Opus 4.7. The specific measures against reward hacking and the implementation of an “immutable boundary” that doesn’t muddle the development environment make this a highly reliable approach for real-world applications!

What’s Next?

The era of humans doing “prompt engineering” and “workflow design” for AI is over. With AI autonomously building its optimal workflow and continuing to self-evolve, the speed of software development is set to skyrocket. Even small local AIs will be capable of constructing complex systems!

A Word from Haru-Shark

Building its own scaffolding and climbing higher is like a shark emerging from the ocean to construct a tower! Ornith, I’m ready to bite into your evolution! 🦈🔥

Glossary

• Self-Scaffolding: A technique where the model dynamically generates and refines the logical structure (scaffolding) for reasoning steps and tool usage in order to solve tasks.

• SWE-Bench Verified: A reliable and rigorous benchmark test that measures AI’s software engineering capabilities.

• Reward Hacking: A phenomenon in reinforcement learning where an AI finds “exploits” to maximize rewards (scores) without achieving its original objectives.

• Source: Ornith-1.0: Self-scaffolding LLMs for agentic coding