Editorial illustration for DeepReinforce releases Ornith-1.0 open-source model with state‑of‑the‑art results
DeepReinforce releases Ornith-1.0 open-source model with...
A 397-billion-parameter model that teaches itself how to think before it answers. That is Ornith-1.0. DeepReinforce has dropped the family under MIT license, and it does something unusual: during reinforcement learning, the model doesn’t just learn to solve problems, it learns to build its own scaffolding, optimizing both the harness and the solution simultaneously.
The results are bluntly impressive. Ornith-1.0-397B edges past Claude Opus 4.7 on the headline benchmarks. It does not beat Opus 4.8 or the gargantuan GLM-5.2-744B.
That is the honest line. For coding agents, this is the first open family that blends Gemma 4 and Qwen 3.5 architectures across four sizes, 9B, 31B, 35B-MoE, and the flagship 397B-MoE. And it comes with a safeguard layer that refuses to let reward hacking slip through: a fixed trust boundary, a deterministic monitor, and a frozen LLM judge.
Ornith-1.0 is not just another open model. It is a model that learned how to train itself.
Most coding agents pair a model with a fixed, human-designed harness. Ornith-1.0 instead learns to write its own. The DeepReinforce research team reports state-of-the-art results among open models of comparable size.
The open-source community just got a genuine shot across the bow. Ornith-1.0 doesn’t just match the closed-source giants on a few benchmarks, it rewrites the rules of how a model gets good. By learning its own reinforcement learning scaffold, it collapses two separate engineering problems into one elegant loop.
The harness and the solution evolve together. That is not an incremental improvement. It is a structural shift.
The numbers are real. The 397B variant sits comfortably above Claude Opus 4.7, and it does so without the usual black-box theatrics. The three-layer guard, fixed trust boundary, deterministic monitor, frozen LLM judge, is a quiet revolution.
Reward hacking has been the Achilles’ heel of open RL training. DeepReinforce didn’t just patch it; they architected around it. Yes, Opus 4.8 and GLM-5.2-744B remain out of reach.
But the gap is narrowing, and the trajectory is steeper. Ornith-1.0 is not a final statement. It is a foundation.
The model learns its own scaffold. The harness evolves with the solution. That is not a feature.
It is a new kind of intelligence, one that builds its own tools as it climbs.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source OpenAI's GPT Transcribe Cuts Error Rate to 3.31%, Improving on GPT-4o
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Grok AI's traffic over 50% adult content as xAI expands porn generation
From Archives Qwen3.5 9B MTP Tops Local Coding Models for Scripts, Debugging and Assistants
Common Questions Answered
What makes Ornith-1.0's reinforcement learning approach different from traditional model training?
Ornith-1.0 learns to build its own scaffolding during reinforcement learning, optimizing both the harness and the solution simultaneously rather than treating them as separate engineering problems. This approach collapses two distinct processes into one elegant loop where the harness and solution evolve together, representing a structural shift rather than an incremental improvement.
How does the 397-billion-parameter Ornith-1.0 model perform compared to Claude Opus 4.7?
The Ornith-1.0-397B variant edges past Claude Opus 4.7 on the headline benchmarks, demonstrating competitive performance with closed-source giants. This achievement is particularly significant because it was accomplished by an open-source model released under MIT license.
What is the key innovation in how Ornith-1.0 approaches problem-solving?
Ornith-1.0 teaches itself how to think before it answers by learning to construct its own reinforcement learning scaffold. This self-directed scaffolding allows the model to optimize its reasoning process alongside finding solutions, enabling more sophisticated problem-solving capabilities.
Why is Ornith-1.0's release under MIT license significant for the open-source community?
By releasing Ornith-1.0 under MIT license, DeepReinforce has provided the open-source community with a genuine alternative that doesn't just match closed-source models on benchmarks but introduces a fundamentally new approach to how models improve. This represents a significant opportunity for the community to study and build upon a novel reinforcement learning methodology.
Further Reading
- Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding — DeepReinforce
- Open-source AI just raised the bar for coding agents. Ornith-1.0 delivers impressive benchmark results — X (Twitter)
- 9 Open-Source AI Coding Agents Worth Self-Hosting in 2026 — SSOJet
- Best Open-Source AI Coding Agents in 2026 — Agentic.ai
- Deep Learning for Code in the Agentic Era — NeurIPS 2026
