Agentica and Together AI have made a significant advance in AI-assisted software development with the release of DeepCoder-14B-Preview. This open-source code reasoning model competes with proprietary counterparts like OpenAI's o3-mini-2025, showcasing 60.6% Pass@1 accuracy on LiveCodeBench. What sets DeepCoder apart is the complete transparency of its dataset, code, and training methods. The team's innovative approach included developing a curated training set of 24,000 coding problems and using the enhanced GRPO+ algorithm to improve training stability and context reasoning, which collectively aims to democratize access to cutting-edge AI tools for developers worldwide.
DeepCoder-14B-Preview from Agentica and Together AI is an open-source code reasoning model that rivals proprietary solutions with fewer parameters.
DeepCoder achieves 60.6% Pass@1 accuracy on LiveCodeBench, matching models with more parameters, showcasing its competitive edge.
The team behind DeepCoder created a high-quality training set and employed GRPO+, a refined algorithm that enhances stability and long-context reasoning.
By releasing all components of DeepCoder openly, the project significantly contributes to democratizing advanced AI coding tools for the global developer community.
Collection
[
|
...
]