Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model
github.com
June 14, 2026
1 min read
🔥🔥🔥🔥🔥
65/100
Summary
Rio-3.5-Open-397B is a 397B model that combines weights from Nex-N2_pro and Qwen3.5-397B-A17B in a ratio of 0.6 to 0.4. There is no evidence of independent training for this model, indicating it is a direct merge of existing models.
Key Takeaways
- The Rio-3.5-Open-397B model is a direct element-wise merge of the Nex model and the Qwen3.5-397B-A17B base, with a ratio of approximately 0.6 Nex to 0.4 Qwen.
- The Rio model identifies itself as "Nex, from Nex-AGI" 79% of the time when its hard-coded prompt is removed.
- All weight tensors in the Rio model consistently reflect the 0.6/0.4 blend of Nex and Qwen across all layers and components.
- There is no evidence of original training for the Rio model, as it replicates the backstory of Nex verbatim.
Coverage & discussions
Hacker News · github.com
Community discussions
Community Sentiment
Positives
- The concept of merging open weight models to create a potentially superior AI is intriguing, suggesting a collaborative approach to model development that could enhance performance.
- The involvement of Rio de Janeiro's IT department in AI work, even with limited resources, indicates a growing interest in leveraging AI technologies for public benefit.
Concerns
- Claims of improved outputs from post-training were misleading, raising serious ethical concerns about transparency and accountability in AI model development.
- The model's actual performance is questionable, as it appears to be a weighted merge rather than a genuine fine-tune, which could mislead users about its capabilities.
- Concerns about the misuse of taxpayer money for potentially fraudulent AI claims highlight the need for stricter oversight in public sector AI initiatives.
