DeepReinforce open-sources Ornith-1.0 coding models

DeepReinforce's RL-driven scaffolding training approach demonstrates how open-source models can achieve SOTA agentic performance without massive parameter scaling. This family could significantly lower the cost of running local code-editing agents.

• –The self-improving RL scaffold generation allows the models to optimize search trajectories, leading to superior bug localization and multi-file refactoring.
• –With models ranging from 9B dense to a massive 397B MoE, developers can run capable agentic models locally on edge devices or host them on high-throughput infrastructure.
• –Achieving 82.4 on SWE-bench verified places the 397B MoE variant among top-tier open-source reasoning models.
• –Native compatibility with vLLM, SGLang, and standard OpenAI-compatible APIs makes drop-in replacement in current agent architectures trivial.
• –Released under an MIT license, it offers a fully permissive alternative for enterprise environments requiring strict licensing compliance.