Model WarsJune 25, 2026via MarkTechPost
DeepReinforce Releases Ornith-1.0: An Open-Source Coding Model Family That Learns Its Own RL Scaffolds
Why it matters
Ornith-1.0 combines novel RL training (learned scaffolds vs. fixed harness) with open-source licensing to challenge closed-model dominance in code generation. This is a capability + distribution play that matters for teams building on OSS foundations.
Key signals
- Ornith-1.0 flagship: 397B parameters
- SWE-Bench Verified score: 82.4
- Built on Gemma 4 and Qwen 3.5 base models
- MIT license (fully open-source weights)
- Novel training approach: learned RL scaffolds (vs. fixed harness)
- Released by DeepReinforce
- Published June 25, 2026
The hook
82.4 on SWE-Bench Verified. DeepReinforce's new open-source coding model just reset the bar—and it learns its own RL scaffolds.
DeepReinforce released Ornith-1.0, an open-source coding model family built on Gemma 4 and Qwen 3.5. Instead of a fixed harness, the model learns its own scaffold during reinforcement learning. The 397B flagship reports 82.4 on SWE-Bench Verified, with all weights under the MIT license.