Skip to main content

Read the full write-up

Notion blog post with full details
 DeepCoder is a 14B coding model that achieves 60.6% Pass@1 on LiveCodeBench, an 8% improvement over the base model. It matches o3-mini and o1 performance on competitive programming benchmarks.

​
Results

ModelParametersLiveCodeBench Pass@1
DeepCoder14B60.6%
o3-mini (Low)Unknown~60%
o1Unknown~60%

​
Approach

Fine-tuned from DeepSeek-R1-Distill-Qwen-14B on coding competition problems using RL with code execution feedback. See the DeepCoder example for instructions on reproducing this with rLLM. Released: April 2025