daVinci-Dev

model
SII
Agent-native mid-training for software engineering. 32B and 72B (Qwen2.5 base). SWE-Bench Verified: 32B 56.1%, 72B 58.5% — surpasses Kimi-Dev using less than half the mid-training tokens (73.1B total). OpenSWE-72B reaches 66.0% (SOTA among SFT methods).
Paper (arXiv)HuggingFace

Model Details

Architecture DENSE
Base model qwen2.5

Variants

Name Parameters Notes
daVinci-Dev-32B — 32B, mid-trained on Qwen2.5
daVinci-Dev-72B — 72B, mid-trained on Qwen2.5

Paper

arXiv: 2601.18418

codingagenticopen-weight