Ornith 1.0: This is new class of self-improving model
Ornith 1: Open-Weight Agentic Coding Models That Write Their Own Harnesses (and Beat Bigger Models) In this video: I break down Ornith 1, a new family of open-weight models built for agentic coding that can outperform much larger models on benchmarks like Terminal Bench, with the 397B model nearing closed-source performance (Opus 4.8). The key idea isn’t just scores—it’s how Ornith is trained to generate both solution rollouts and a task-specific harness (memory, retries, error handling) in a single loop, using reinforcement learning (GRPO) so rewards update both the solution and the scaffold. I cover reward-hacking risks and the three-layer defenses: locked boundaries, deterministic monitoring, and a frozen judge model. I also share my own Ollama tests on an M2 Max comparing Qwen 3.5 9B base vs Ornith 1 9B (8-bit): similar accuracy, but Ornith is ~3× cheaper (up to 20× on some tasks), while long-horizon “honesty under pressure” seems to require larger scale (35B+). LINKS: https://deep-reinforce.com/ornith_1_0.html https://huggingface.co/collections/deepreinforce-ai/ornith-10 My voice to text App: whryte.com Website: https://engineerprompt.ai/ RAG Beyond Basics Course: https://prompt-s-site.thinkific.com/courses/rag Signup for Newsletter, localgpt: https://tally.so/r/3y9bb0 Let's Connect: 🦾 Discord: https://discord.com/invite/t4eYQRUcXB ☕ Buy me a Coffee: https://ko-fi.com/promptengineering |🔴 Patreon: https://www.patreon.com/PromptEngineering 💼Consulting: https://calendly.com/engineerprompt/consulting-call 📧 Business Contact: engineerprompt@gmail.com Become Member: http://tinyurl.com/y5h28s6h 💻 Pre-configured localGPT VM: https://bit.ly/localGPT (use Code: PromptEngineering for 50% off). Signup for Newsletter, localgpt: https://tally.so/r/3y9bb0 00:00 Ornith Models Overview 02:46 Self Written Harnesses 04:31 Reward Hacking Risks 05:30 Three Layer Defenses 06:17 My Ollama Test Setup 07:40 Private Bench Results 08:33 Long Horizon Honesty Test 09:53 Key Takeaways on 9B 10:42 Caveats
