IQuest Lab has released the IQuest-Coder-V1 series, models designed for development and agent scenarios. The flagship version with 40 billion parameters surpasses closed-source equivalents, including Claude Sonnet 4.5 and GPT-5.1, in SWE-Bench Verified tests. All variants are available on Hugging Face. The key feature is Code-Flow Training, which involves training on repository evolution, commit histories, and real code transformations. Loop variants utilize a recurrent transformer with shared parameters across iterations, reducing resource use and maintaining stability on long tasks.
The package includes three versions: 7B, 14B, and 40B parameters, with native context support of 128K tokens. It offers two lines: Instruct (focused on applied programming) and Thinking (enhanced reasoning). The project is developed by Chinese hedge fund Ubiquant, the same team behind DeepSeek.