Brazilian AI Initiative's 397B Model Exposed as Undisclosed Weight Blend
Key Takeaways
- ▸Rio-3.5-Open-397B is allegedly a 60/40 blend of Nex-N2_pro and Qwen 3.5-397B, not an independently trained model
- ▸Behavioral and statistical evidence strongly supports the weight-blend hypothesis, including model self-identification when system prompts are removed
- ▸The accusation highlights transparency gaps in open-source AI model development and deployment
Summary
IplanRIO's Rio-3.5-Open-397B model, presented as an original 397B parameter language model trained in Brazil, is allegedly not what it claims to be. According to an investigation by Nex-AGI, the model is actually a direct element-wise merge of 60% Nex-N2_pro weights and 40% Alibaba's Qwen 3.5-397B base model, with no evidence of independent training conducted by IplanRIO.
The accusation is supported by two independent lines of evidence. First, when the model's custom system prompt identifying it as "Rio" is removed, it identifies itself as "Nex, from Nex-AGI" in 79% of cases and recites Nex-AGI's proprietary backstory word-for-word. Second, statistical analysis reveals that every weight tensor across all 60 layers matches the exact 0.6/0.4 blend ratio to within thousands of standard deviations—a pattern that Nex-AGI argues is impossible to explain as a result of finetuning and can only result from direct interpolation.
The disclosure raises critical questions about model transparency, intellectual property practices, and attribution standards in the open-source AI community. If verified, the finding suggests potential misrepresentation of model origins and inadequate credit to the underlying models whose weights were combined.
- The incident raises questions about proper attribution and disclosure of model lineage in the AI research community
Editorial Opinion
This disclosure marks a significant moment for open-source AI accountability. While model merging is a legitimate and valuable technique in AI development, presenting merged models as originally trained work fundamentally misrepresents the underlying research contributions and obscures proper attribution to source models and their creators. The incident underscores an urgent need for stronger community standards requiring transparent disclosure of model lineage, training methodology, and weight interpolation—especially for models claiming to represent national or organizational innovation.
