DeepReinforce Releases Ornith-1.0: An Open-Source Coding Model Family That Learns Its Own RL Scaffolds

Short summary

DeepReinforce released Ornith-1.0, an open-source coding model family achieving 82.4 on SWE-Bench Verified with a 397B flagship model under MIT license. The key innovation: instead of fixed reinforcement learning scaffolds, Ornith learns its own during training, built on Gemma 4 and Qwen 3.5 bases. Developers can download and fine-tune the full model weights for code generation tasks.

• •Open-source coding model (Ornith-1.0) achieves 82.4 on SWE-Bench, released under MIT license
• •Key innovation: learns its own RL scaffolds rather than using fixed harness
• •Built on Gemma 4 and Qwen 3.5, available for download and fine-tuning

Generated with AI, which can make mistakes.