TL;DR
Rio de Janeiro’s claimed ‘homegrown’ large language model appears to be a blend of existing models Nex and Qwen, not an independently trained system. Evidence shows the model’s weights are a direct interpolation of these two models, raising concerns about its originality.
Recent technical analysis reveals that Rio de Janeiro’s ‘homegrown’ large language model, officially presented as an original 397 billion parameter system, is in fact a direct blend of existing models Nex and Qwen3.5-397B, not a newly trained, independent model.
According to findings shared on Hacker News, the so-called Rio 3.5-Open-397B model’s weights are an element-wise merge of the Nex model and the official Qwen3.5-397B base. The analysis shows that approximately 60% of the model’s weights correspond to Nex, with the remaining 40% matching Qwen. This was confirmed by testing the model’s responses and examining its weight tensors, which align with this interpolation across all layers.
Further, when the model’s system prompt was removed, it identified itself as ‘Nex, from Nex-AGI’ 79% of the time, and as ‘Rio’ only 0%, reciting the Nex backstory verbatim. No evidence suggests any independent training or fine-tuning beyond this interpolation, raising questions about the claim of a fully original model.
Implications for Transparency in AI Development
This development matters because it highlights transparency issues in AI claims—specifically, whether local or government-backed models are genuinely developed in-house or are composites of existing models. Such revelations could impact public trust, regulatory scrutiny, and the perceived originality of AI initiatives in Brazil and beyond.
Advanced Language Tool Kit: Teaching the Structure of the English Language
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Background on Rio’s AI Claims and Model Origins
Rio de Janeiro announced its ‘homegrown’ LLM, claiming it was trained independently by local researchers using proprietary data. However, recent technical evidence suggests the model is a weighted blend of Nex and Qwen models, which are publicly available. This follows a broader pattern where AI models are sometimes presented as original when they are composites or fine-tuned versions of existing systems. The analysis was prompted by community scrutiny on Hacker News, where users examined the model’s weights and responses.
“The weights of Rio’s model are a near-perfect element-wise blend of Nex and Qwen, with no signs of independent training or fine-tuning.”
— an anonymous researcher on Hacker News
Extent of Original Development Still Unclear
While the weight analysis strongly suggests the model is a blend of Nex and Qwen, it remains unclear whether any additional fine-tuning or proprietary training was conducted beyond the interpolation. The official claims of a fully original model have not been substantiated, but definitive proof of all development processes is not yet available.
Further Technical Analyses and Official Clarifications Pending
Researchers and industry observers are expected to conduct more detailed examinations of the model’s training data and procedures. Official statements from the Rio de Janeiro AI team could clarify whether any additional training occurred. Monitoring these developments will help determine if the model’s origin aligns with the official narrative or remains a composite system.
Key Questions
Is Rio’s model truly ‘homegrown’?
Current evidence suggests the model is a blend of Nex and Qwen, not a fully independently trained system, despite official claims.
What does this mean for AI transparency in Brazil?
This raises questions about transparency and honesty in AI development claims, especially for government-backed projects.
Could the model still have been fine-tuned or trained further?
It is possible, but no evidence currently supports additional proprietary training beyond the weight interpolation observed.
Will there be official clarification from Rio authorities?
It is not yet clear if Rio officials will respond or provide further details about the model’s development process.
How does this affect the perception of local AI initiatives?
This incident may impact trust in local AI projects if claims of originality are proven false or misleading.
Source: Hacker News
