DeepReinforce releases Ornith-1.0 open-source coding model family with self-scaffolding RL
Key Questions
What makes Ornith-1.0 different from other coding models?
Ornith-1.0 learns its own RL scaffolds during training, reducing the need for manual agent design. This addresses a key pain point in building effective coding agents.
What are the hardware requirements for Ornith-1.0 models?
The 35B MoE variant fits within 32-64GB VRAM, making it accessible for many local setups. Smaller 9B-35B sizes are particularly suited for resource-constrained environments.
How does Ornith-1.0 perform on coding benchmarks?
It achieves 82.4 on SWE-bench Verified, outperforming larger models like Qwen3.5-397B. Performance is strong relative to other open models of similar scale.
Ornith-1.0 (9B-35B, MIT) learns its own RL scaffolds, addressing agent design pain points. 35B MoE variant fits 32-64GB VRAM. Achieves strong benchmarks vs comparable open models, though not frontier. Smaller sizes highly relevant for local agent deployment.