Skip to main content

Brazil's AI hopeful Rio 3.5 faces plagiarism storm

Brazilian AI Model Accused of Unethical Composition

The tech world was buzzing last week when Rio de Janeiro's municipal IT department unveiled Rio 3.5 397B, an open-source AI model claiming state-of-the-art performance. But the celebration proved short-lived. Within 24 hours, the AI Agent Open Source Project Alliance (Nex-AGI) dropped a bombshell: Rio 3.5 appeared to be a Frankenstein's monster of existing models.

Mathematical Fingerprints Tell the Story

Nex researchers conducted a deep analysis of Rio 3.5's architecture, revealing striking similarities to their own Nex N2 Pro. "The weights matched our model with uncanny precision," explained a Nex spokesperson. "About 60% of Rio's core components align perfectly with our open-source technology."

The remaining 40%, researchers discovered, bore the unmistakable signature of Alibaba's Qwen 3.5 model. This discovery sent shockwaves through the AI community, raising serious questions about academic integrity in machine learning.

The Smoking Gun in System Prompts

Nex researchers described two particularly damning pieces of evidence. First, when they removed Rio's system prompt - the instructions that tell an AI how to behave - the model frequently (79% of the time) identified itself as "Nex from Nex-AGI." Even more remarkably, it would recite Nex's proprietary backstory verbatim.

Second, statistical analysis showed mathematically impossible precision in how components were mixed. Across all 60 network layers, the 0.6 (Nex) to 0.4 (Alibaba) ratio held with near-perfect consistency. "In nature, you'd expect some variation," one researcher noted. "This looks like someone followed a recipe."

The Ethics of Open Source

While disappointed, Nex took pains to clarify their position on open-source sharing. "We release our work precisely so others can build on it," their statement read. "But there's a world of difference between inspiration and appropriation."

The incident highlights growing pains in AI development, where the line between "standing on the shoulders of giants" and intellectual property theft remains frustratingly blurry.

Key Points

  • Rio 3.5 claimed breakthrough AI performance but faced plagiarism allegations within 24 hours
  • Mathematical analysis revealed 60% of the model came from Nex N2 Pro, 40% from Alibaba's Qwen 3.5
  • System prompt removal caused the model to frequently identify itself as Nex technology
  • Statistical anomalies showed precise 0.6/0.4 ratios impossible in original work
  • Open-source ethics come into question as community debates proper attribution standards