Together AI releases DeepCoder-14B-Preview: Open-source 14B code reasoning model
Action Required
Developers can now leverage a competitive code reasoning model without proprietary licensing costs, potentially accelerating software development and research.
AI Impact Summary
Together AI has released DeepCoder-14B-Preview, a fully open-source 14B parameter code reasoning model achieving 60.6% Pass@1 accuracy on LiveCodeBench, matching o3-mini. This release leverages reinforcement learning with a curated dataset of 24K verified coding problems, including TACO Verified, PrimeIntellect SYNTHETIC-1, and LiveCodeBench data. The team has also open-sourced key optimizations like verl-pipe, accelerating training by 2x, and details the training recipe including GRPO+ with No Entropy Loss, No KL Loss, Overlong Filtering, and Clip High, enabling the model to generalize to 64K context length.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- high