GAIR/daVinci-Dev-72B

Overview of daVinci-Dev-72B

daVinci-Dev-72B is a 72.7 billion parameter model from the GAIR family, specialized in agentic software engineering. It introduces a novel "agent-native mid-training" approach to bridge the gap between static pretraining data and the dynamic, feedback-rich environments encountered by real code agents. The training incorporates two distinct trajectory types: contextually-native trajectories derived from GitHub Pull Requests (PRs) and environmentally-native trajectories collected from executable repositories with genuine tool and test outputs.

Key Capabilities

• Agentic Software Engineering: Designed to operate effectively within agentic scaffolds, such as SWE-Agent, for automated software development tasks.
• High Performance on SWE-Bench: Achieves a 58.5% Pass@1 on SWE-Bench Verified, demonstrating state-of-the-art performance among open training recipes for its size, despite originating from a non-coder base model (Qwen2.5-Base).
• Generalization: Shows improved performance on standard code benchmarks like HumanEval/EvalPlus and scientific reasoning benchmarks such as GPQA/SciBench.
• Specialized Training Data: Utilizes 68.6 billion tokens of contextually-native PR trajectories and 4.5 billion effective tokens of environmentally-native executable trajectories, including both test-passing and non-passing rollouts.

Good for

• Automated Code Generation and Debugging: Ideal for applications requiring an LLM to interact with development environments, propose code changes, and validate solutions.
• Software Agent Development: Provides a robust foundation for building and deploying autonomous software engineering agents.
• Research in Agentic LLMs: Useful for researchers exploring advanced training methodologies for LLMs in dynamic, feedback-driven environments.