Introduction to DeepCoder-14B-Preview
DeepCoder-14B-Preview is an open source code generation model developed by the Agentica team and released on the Hugging Face platform. It is optimized based on DeepSeek-R1-Distilled-Qwen-14B and uses distributed reinforcement learning (RL) techniques. The main goal of this model is to provide developers with efficient code generation capabilities that are particularly suitable for programming competitions and large-scale project development.
Main features
- Raw Growth Code: supports contexts up to 64K tokens, generates and handles very long code
- Highly accurate output: 60.6% Pass@1 accuracy in LiveCodeBench v5 test
- Completely open source: Provides complete model weights, training data and scripts
- Multi-Programming Task Support: for competition question answering, code debugging and project development
- Optimized Long Context Reasoning: Ensuring long code generation capabilities through GRPO+ and DAPO technologies
With 14 billion parameters and performance close to OpenAI's o3-mini, this model is the current leader among open source code generation models.
This answer comes from the articleDeepCoder-14B-Preview: an open source model that specializes in code generationThe
May not be reproduced without permission:AI productivity tools " What is DeepCoder-14B-Preview and what are its key features?
