chinese-llm-benchmark

Why we included this project

Choosing a Chinese-language model for a real product usually means comparing vague marketing claims, and ReLE tries to replace that with a fine-grained look at what models actually do. The project runs hundreds of commercial and open-weight models through roughly seven domains and about 300 sub-areas, covering school subjects, medical licensing, finance, law, and agent tool use, then publishes the scores as live leaderboards. The more unusual part is the defect library with over two million recorded failures, which lets researchers and engineers hunt down specific weaknesses instead of trusting an aggregate score. If you're picking a model for a Chinese-facing product or sanity-checking a private model before deployment, the per-dimension results give you something more actionable than a single rank. And because the benchmark keeps adding newly released models, the comparisons stay relevant as the field moves.