DeepCoder-14B-Preview Open-Source 14B Code-Reasoning Model Reaches o3-mini-Level Performance

AI Impact Summary

DeepCoder-14B-Preview is a RL-finetuned 14B code-reasoning model that achieves 60.6% Pass@1 on LiveCodeBench, matching o3-mini's performance with a smaller parameter footprint. The release includes open-source dataset, code, training logs, and verl-pipe, plus infrastructure notes (100+ sandboxes, 32K to 64K context) and a detailed RL training approach, enabling reproducibility and on-prem or cloud deployment. This open, capable coder democratizes access to competitive coding models, potentially shifting cost-and-feature expectations for enterprise tooling and accelerating internal development of code-generation features.