GAIR/daVinci-Dev-32B

TEXT GENERATIONConcurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jan 25, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

GAIR/daVinci-Dev-32B is a 32.8 billion parameter language model from the GAIR family, specifically trained for agentic software engineering. It utilizes agent-native mid-training on PR-derived and environmentally-native trajectories to reduce the distribution mismatch between static pretraining and dynamic coding environments. This model excels at complex software development tasks, achieving 56.1% on SWE-Bench Verified, and is built upon the Qwen2.5-Base architecture.

Loading preview...

Overview of daVinci-Dev-32B

daVinci-Dev-32B is a 32.8 billion parameter model from the daVinci-Dev family, developed by GAIR for agentic software engineering. This model introduces agent-native mid-training to bridge the gap between static pretraining data and the dynamic, feedback-rich environments encountered by real code agents. It leverages 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: Specifically optimized for tasks requiring an agent-like approach to code development, including understanding context, proposing edits, and validating with tests.
  • High SWE-Bench Performance: Achieves a 56.1% Pass@1 on the challenging SWE-Bench Verified benchmark, demonstrating strong performance in automated software development tasks.
  • Generalization: Shows improvements on standard code benchmarks like HumanEval/EvalPlus and scientific reasoning benchmarks, indicating broader applicability beyond its core specialization.
  • Robust Training Methodology: Built on the Qwen2.5 base model, it undergoes agent-native mid-training on a massive dataset (68.6B PR-derived tokens and 4.5B effective environmentally-native tokens), followed by Supervised Fine-Tuning (SFT) on test-passing trajectories.

Good For

  • Automated Code Generation and Refactoring: Ideal for systems that need to autonomously understand, modify, and validate code.
  • Integration with Agent Frameworks: Designed to be used effectively within agent scaffolds like SWE-Agent.
  • Research in AI for Software Engineering: Provides a strong baseline and methodology for further exploration into agentic LLMs for development tasks.