Home / Technology / AI Coder Trained in 4 Days Outperforms Giants
AI Coder Trained in 4 Days Outperforms Giants
8 Jan
Summary
- NousCoder-14B, an open-source AI, was trained in just four days.
- It matches or surpasses larger proprietary AI coding systems.
- The model achieved 67.87 percent accuracy on LiveCodeBench v6.
Nous Research has released NousCoder-14B, an open-source AI model capable of competitive programming, trained in just four days using advanced GPUs. This model achieves a 67.87 percent accuracy rate on LiveCodeBench v6, demonstrating performance comparable to or exceeding larger, proprietary AI systems. The rapid development highlights the accelerating pace of AI in software development.
The radical openness of NousCoder-14B is a key differentiator. Nous Research has made model weights, the reinforcement learning environment, and training harness publicly available, allowing for reproducibility and further research. The training process utilized a "verifiable rewards" system, with solutions executed and validated against numerous test cases.
This breakthrough comes amid intense competition, including Anthropic's Claude Code. Nous Research's commitment to open-source alternatives and transparency suggests a growing trend towards community-driven AI development. The model is available on Hugging Face under an Apache 2.0 license.
