Preprint: 8B Small Model Outperforms Frontier LLMs in GPU Kernel Optimization

An 8-billion parameter language model named "Little Dragon" has achieved higher per-attempt correctness (91.27% for construction, 92.01% for refinement) and better kernel performance than six frontier large language models, including GPT-6 Astra and Opus 5, in GPU kernel optimization tasks. These results come from an arXiv preprint submitted on October 14, titled "How to Train Your Little Dragon."

The research team used the TVM MetaSchedule framework to build training data on 54,116 distinct workloads. By combining supervised fine-tuning with reinforcement learning, they taught the small model to generate complete compiler optimization scripts directly, rather than just ranking candidates or generating raw code. This compact, lossless representation allows deterministic decoding to reconstruct exact compiler traces.

This finding suggests that small models specifically trained for narrow domains like compiler optimization may outperform general-purpose giant models in both efficiency and effectiveness, offering a new path to reduce inference costs in AI infrastructure. However, these are author-reported benchmarks and have not yet been independently reproduced.