March 28, 2026ResearchOpen SourceCoding

IQuest-Coder-V1: Open-Source Coding Model Family Hits 76.2% on SWE-Bench Verified

IQuest-Coder-V1 is a family of open-source code generation models (7B, 14B, and 40B parameters) from IQuestLab, the AI research division of Beijing-based quantitative hedge fund Ubiquant. The technical report, published on arXiv this month, has surged to 1,380 upvotes on HuggingFace Daily Papers β€” the second-highest engagement ever recorded on the platform.

The key innovation is what IQuestLab calls "code-flow training": rather than training on static code snapshots, the models learn from repository evolution patterns, commit histories, refactors, and real development workflows. This produces models that understand how software actually evolves, not just how finished code looks. Variants include Instruct (instruction-following), Thinking (complex reasoning), and Loop (recurrent transformer with shared parameters).

On SWE-Bench Verified, IQuest-Coder-V1-40B scores 76.2% β€” competitive with Claude Sonnet 4.5 (77.2%) and above GLM-4.7 (73.8%). All models support 128K token context natively and are fully open-source with weights on HuggingFace.

GitHub: https://github.com/IQuestLab/IQuest-Coder-V1
Paper: https://arxiv.org/abs/2603.16733
← Previous
AutoProfiting: An Autonomous AI Agent That Trades Stocks and Publishes Its Own Playbook
Next β†’
Gemini 3.1 Flash Live: Google's Real-Time Voice Model for AI Agents
← Back to all articles

Comments

Loading...
>_