NextCoder

NextCoder is a series of code-editing language models adapted from QwenCoder-2.5 using a novel fine-tuning method called Selective Knowledge Transfer (SeleKT). SeleKT performs a dense gradient step to identify edit-critical weights and then projects them sparsely back onto the base model, enabling targeted adaptation without full retraining. The series ships in 3B, 7B, 14B, and 32B parameter variants and is available on Azure AI Foundry and Hugging Face.

NextCoder outperforms peer code-editing models across five benchmarks covering Aider polyglot editing, code generation, and code repair. On Aider polyglot, NextCoder-32B reaches 88.9% accuracy, matching or exceeding standard supervised fine-tuning approaches on multiple metrics while preserving the strengths of the base QwenCoder-2.5 model. The result is a practical demonstration that targeted, surgical adaptation of code models — rather than expensive full retraining — is a viable path to specialized code-editing capability, and it slots into Microsoft’s broader stack of developer-tools models alongside BugPilot and Debug-gym.