IQuest Coder – code LLMs for software engineering and competitive programming (iquestlab.github.io)

IQuest-Coder, an open-source AI model, has achieved a remarkable 81.4% accuracy on SWEBench, indicating significant advancements in coding benchmarks. This model incorporates innovative features such as Multi-stage Code-Flow training to capture code evolution, an expanded reasoning dataset for enhanced long-context reasoning, and a unique Loop architecture that not only reduces memory overhead but also increases throughput. These improvements allow IQuest-Coder to achieve performance levels comparable to hundreds-of-billions-parameter MoE models, all while minimizing training costs and maximizing deployment efficiency on consumer-grade GPUs. The significance of IQuest-Coder for the AI/ML community lies in its ability to handle real-world tasks with greater reliability, thanks to an intricate training pipeline that includes Pre-Train, Annealing, Mid-Train, and Post-Train stages. The dual post-training paths—focused on reasoning and general assistance—further enhance its usability. As the model demonstrates leading results on key coding benchmarks, it reshapes the performance capabilities of language models, optimizing them for scalable and efficient deployment. This represents a crucial step towards advancing AI's role in software development and coding assistance.