Rio de Janeiro Built an AI Model That Beat DeepSeek—But Was Based on Someone Else's Work

Rio de Janeiro, a Brazilian AI company, announced on May 2, 2024, the release of a new large language model (LLM) named "Rio-7B" that purportedly outperformed Alibaba's Qwen1.5-7B-Chat model on several benchmarks. The company claimed Rio-7B achieved a score of 72.5 on the MT-Bench leaderboard, surpassing Qwen1.5-7B-Chat's score of 71.7. However, shortly after the announcement, the AI startup Nex, founded by former Google AI researchers, presented evidence suggesting that Rio-7B was largely based on their proprietary "Nex-7B" model. Nex provided technical details and benchmark comparisons to support their claim, indicating that Rio-7B was essentially a fine-tuned version of Nex-7B, with minimal original contributions from Rio de Janeiro. The allegations, if true, would mean Rio de Janeiro's claimed performance gains were not independently achieved. This situation highlights the ongoing challenges in AI development regarding intellectual property and transparent reporting of model origins and performance.