The city of Rio de Janeiro’s IT agency, IplanRIO, recently launched what it described as a “frontier-class” AI model, Rio 3.5 Open 397B. The model, boasting 397 billion parameters and a permissive open-source license, was presented as a significant achievement by a municipal government in the Global South. It incorporated a novel reasoning layer called SwiReasoning, designed to enhance performance by dynamically switching between explicit and latent reasoning modes based on model confidence. Initially, the model card indicated it was a post-training of Alibaba’s Qwen 3.5 397B, with the added SwiReasoning layer, and reportedly had a development cost significantly lower than comparable commercial systems.
Key Takeaways
- IplanRIO released Rio 3.5 Open 397B, claiming it was a government-developed frontier AI model that outperformed existing benchmarks.
- AI company Nex provided evidence, including mathematical proofs and behavioral analysis, demonstrating Rio 3.5 was a weight merge of their Nex N2 Pro model and Qwen 3.5.
- IplanRIO updated the model’s documentation to acknowledge Nex’s contribution and attributed the original claims to an “incorrect upload.”
- The incident highlights the importance of transparency and proper attribution in the open-source AI community, even for publicly funded projects.
- Model merging is a common and legal practice in AI development, but failing to disclose the foundational models used can lead to reputational damage.
The initial announcement garnered considerable attention, especially after benchmark results were published showing Rio 3.5 surpassing leading models like Qwen 3.7 Plus and DeepSeek v4 Pro on various tasks, including autonomous command execution and mathematical reasoning. This was amplified by social media posts from the Mayor of Rio de Janeiro, celebrating the city’s supposed AI advancement.
However, the narrative shifted dramatically when Nex-AGI, an open-source AI alliance, published findings suggesting that Rio 3.5 was not an independently developed model but rather a derivative work. Nex presented evidence indicating that Rio 3.5 was a weighted merge, approximately 60% Nex N2 Pro and 40% Qwen 3.5. Their analysis included a mathematical proof showing a high degree of collinearity between the weight tensors of Rio 3.5 and their own model, suggesting a direct blend rather than independent training.
Further evidence presented by Nex involved removing Rio 3.5’s system prompt, which led the model to identify itself as “Nex, from Nex-AGI” over 79% of the time. The model also reportedly recited Nex’s specific backstory verbatim. This behavioral analysis, combined with the mathematical weight analysis, strongly suggested that the core of Rio 3.5 was built upon Nex’s existing work.
Long-Term Impact on Blockchain and AI Innovation
This incident, while centered on AI model development, has implications that resonate within the broader blockchain and Web3 ecosystem. The core tension here is between open innovation, a foundational principle of blockchain, and the presentation of derivative work as novel. In blockchain, transparency and verifiable provenance are paramount. Layer 2 solutions, for instance, build upon the security of Layer 1 networks, and their value is often tied to how effectively they leverage and extend that base layer’s capabilities. Similarly, in Web3 development, projects often rely on open-source protocols and smart contracts. Misrepresenting the origin of foundational components, as seen in the Rio 3.5 case, undermines trust – a critical currency in decentralized systems.
The emphasis on open-source models and the subsequent scrutiny highlights a maturing ecosystem where community-driven verification acts as a powerful governance mechanism. This mirrors the ethos of decentralized autonomous organizations (DAOs) and open-source blockchain projects, where community consensus and transparency are essential for growth and security. Furthermore, the low cost reported for Rio 3.5’s development, even as a merge, points towards efficiency gains that can be achieved by leveraging existing AI models, a concept analogous to how Layer 2 solutions reduce transaction costs on blockchains. As AI becomes increasingly integrated into Web3 applications, the principles of open-source collaboration, clear attribution, and verifiable contributions demonstrated by entities like Nex will become even more critical for fostering sustainable innovation and trust.
Following Nex’s revelations, IplanRIO updated the model card on Hugging Face. The benchmark performance claims were removed, and Nex was credited as a source model. The agency explained the situation as an “incorrect upload,” stating that the raw merged version was released instead of a final, distilled model that would have incorporated further independent work. This response acknowledges the merge but suggests an intention to perform additional development steps. The incident underscores the community’s role in verifying AI developments and the ethical imperative for clear attribution, even when building upon publicly licensed models. The open-source community, accustomed to strict norms around crediting foundational work, quickly identified the discrepancy, emphasizing that while model merging is legal and common, transparency is key.
Original article : decrypt.co