daVinci-Dev: Agent-native Mid-training for Software Engineering

While the emerging field of agentic software engineering has spurred extensive research into post-training, this paradigm alone does not fully address the distribution mismatch between traditional static pre-training and dynamic deployment environments. In this paper, we instead investigate agentic mid-training as a scalable complementary approach. Central to our approach is agent-native data comprising two complementary components: contextually-native trajectories that preserve the complete information flow an agent experiences, offering broad coverage and diversity; and environmentally-native trajectories whose observations stem from actual tool invocations and test executions, providing interaction authenticity. On SWE-Bench Verified, our recipe outperforms the previous open software engineering mid-training recipe Kimi-Dev under two post-training settings with the same base model and agentic scaffold, while using fewer than half mid-training tokens (73.1B). Furthermore, our 32B and 72B models achieve state-of-the-art resolution rates of 56.1\% and 58.5\% among open agentic recipes using agentic scaffolds, despite starting from non-coder Qwen2.5 base models. We also observe performance gains on general code generation and scientific benchmarks. We open-source a significant portion of our datasets, recipes, and model checkpoints to facilitate further research.

Modern AI "coding agents" tackle real software bugs by exploring a codebase, editing files, running tests, and revising when things break — not just writing isolated functions. But the data we usually train them on shows only the finished product: merged code, not how a developer arrived at it. Agents are then forced to learn the back-and-forth of real software work during an expensive final fine-tuning stage. We introduce "agent-native data," training material that mirrors how an agent actually works. We mined millions of public GitHub pull requests to reconstruct end-to-end development stories — which files were examined, what changed, and why — and we let a strong agent run in real software projects to record what actually happens when tests fail and code must be revised. Mixing both sources into an earlier "mid-training" stage gives the model a foundation for agent-like behavior before any task-specific tuning. Our daVinci-Dev-32B and 72B models set state-of-the-art results among open coding recipes on SWE-Bench Verified, despite starting from general-purpose base models and using less than half the mid-training data of the previous best open recipe. We release the data, training recipe, and model checkpoints so others can build on this foundation.